14 December 2020 - Zoom workshop

Track 3 - Smartphone (off-site)

Team Name

Next-Newbie Reckoners

Reference person

Seanglidet Yean

Affiliation
Singtel Cognitive and Artificial Intelligence Lab for Enterprises at NTU (SCALE@NTU); Nanyang Technological University (SG)

Description
We have adopted grid-based approach to train the Random Forest models with an end-to-end pipleline in order to predict the building, floor and location grid using the WiFi Received Signal Strength. In addition, we are working on the Deep Neural Network models as well as feature selection methods to further improve the accuracy.
In this competition, our main objective is to incorporate multi-source of information such as Bluetooth, Smartphone sensor-fusion techniques (e.g. pedestrian dead reckoning) to enhance our method.

Team Name

WHU-five

Reference person

Li Wei

Affiliation
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (LIESMARS), WuHan University (CN)

Description
Training: According to the training data set, establish a geomagnetic fingerprint library, a WIFI fingerprint library, and a lighting library.
Test: Estimate the initial position through WIFI fingerprint positioning, lights, geomagnetism, etc., and then merge the results of PDR, geomagnetism, and wifi positioning to get the final position.
The floor can be judged roughly based on the barometer. When there is a weak change in air pressure between floors, the floor information can be detected by fusing other signals (wifi, geomagnetism, lights, etc.).

Team Name

YAI

Reference person
ChihChieh Yu

Affiliation
Department of Electrical Engineering, Yuan Ze University and National Ilan University (TW)

Description
Our fingerprinting is based on Wi-Fi signal strength (RSS) values expressed in dBm. In this competition, we propose a fuzzy-based pre-processing method so that the RSS entries in the fingerprinting database can be converted into the corresponding defuzzification values. The fuzzy-based pre-processing utilizes five membership functions on the received RSS signals and the parameters associated with these membership functions are trained by the training dataset. Finally, we obtain a non-linear mapping function, which mapping the RSS values to the real values between the range of 0 to 4 as shown in Fig. 1. More specifically, we adopt air pressure value for floor judgment, acceleration value for step detection, and Wi-Fi signal for fingerprinting table positioning.

Team Name
Indora

Reference person
Miroslav Opiela

Affiliation
Institute of Computer Science, P.J. Šafárik University, Faculty of Science, Jesenná 5, 041 54 Košice (SK)

Description

Pedestrian dead-reckoning, map model and bayesian filtering are essential components of the proposed localization system. The main research focus is on a low-dimensional grid-based bayesian filtering, as a less elaborated alternative to Kalman and Particle filters widely used for the positioning. A semi-automatically generated map model helps to reduce the localization error introduced by noisy sensor measurements and inaccurate system configuration, e.g., a step length estimation. The floor transition detection based on barometer measurements allows the system to estimate the user position within a single floor. Step length estimation is incorporated in the particle filter. The light sensor improves the ability to distinguish between two similar corridors and detect indoor-outdoor transitions.

Team Name
UMinho

Reference person
Ivo Silva

Affiliation

University of Minho (PT)

Description

The solution is based on a Wi-Fi radio map built with the signals available in the training dataset. It uses sensor fusion to merge the results obtained from the Wi-Fi fingerprinting with data from other sensors, including accelerometers records to detect the person steps, the mobile phone 3D orientation records for heading estimation and pressure sensors for floor transition detection.

Team Name
imec-WAVES/Ghent University

Reference person
Cedric De Cock

Affiliation

imec-WAVES/Ghent University (BE)

Description

Our system consists of 4 parts. The first part uses intertial sensordata to detect steps, estimate step length and heading estimation and to recognize walking modes. The second part uses the PDR algorithm to interpolate between the sparse ground truth positions. The interpolated positions are used to create a radiomap. The third part uses the barometer and RSS measurements from the training data to detect floor transitions, to estimate during which steps the user is using the stairs and to calculate the most likely sequence of floors visited by the user. In the final part, a backtracking particle filter (BPF) is used to estimate the trajectory. The step lengths and headings are used to propagate the particles. The positions of walls are used to remove particles that were propagated through a wall. The weighted centroid of the k best RSS matches is used to weigh each particle, based on its distance to the matched position. The locations of staircases are used to decrease the weight of particles outside of staircases when the barometer detects a height change. The floor transitions are used to load the next floor map and radiomap.

Team Name
TJU

Team members
Liqiang Zhang, Boxuan Chen, Hu Li, Yazhen liao, Qingyuan Gong and Yu Liu

Affiliation

School of Microelectronics, Tianjin University (CN)

Description

A smartphone integrates rich sensors, including inertial sensor, magnetic sensor, light sensor, sound sensor and proximity sensor, etc. Therefore, it provides a possibility to localize pedestrians, which is valuable and challenging. In this competition, step-length and heading system (SHS) is used as our fundamental system. Acceleration information from inertial sensor is used to detect steps. Then the step length can be computed. To our best knowledge, step-length computing is accurate when a pedestrian moves forward on a level ground. However, the estimated step length becomes inaccurate when a pedestrian moves on a slope, such as walking upstairs or downstairs. To address the problem, we propose adaptive step-length estimation method aided by learning-based motion classifier. Since the performance of the SHS degrades with time and walking distance, reference positions should be introduced to modify the positioning error of the system. Therefore, we establish magnetic field and Wi-Fi-based fingerprint databases with the training and validation data. Then a neural network (NN) model is trained to predict the pre-given ground truth positions using geomagnetic information and Wi-Fi received signal strength. Finally, Kalman filter is leveraged to fuse the positioning results from the NN model and the SHS.

Team Name
MAD PDR (Magnetic Assisted Deep PDR)

Reference person
Leonid Antsfeld

Affiliation

Naver Labs Europe (FR)

Description

We have used a stack of complementary localization components that were intelligently fused together to provide a final actor’s itinerary.
The main components are: floor detection, user activity recognition, speed/step length estimation, PDR, WiFi based localization and, finally, a novel component of magnetic field based localization.
Floor detection was done using classic classification methods based on WiFi and barometer data. We have applied spectral analysis on the accelerometer data in order to identify user's activity (walking, standing, going up or down the stairs). Next, we have used peaks detection of accelerometer data in order to detect steps. We extract acceleration features in order to learn the step length / speed in a given sliding window. Together with the orientation sensor it allowed us to build a PDR which was a first order approximation of the itinerary.
The PDR is known for being accurate locally but drifting over time. In order to compensate for this drift, we have used global localization techniques based on WiFi VAE that was presented at IPIN’19 and recently developed LSTM based magnetic field localization algorithm.
Finally, the itinerary was adjusted using a map that was extracted from the training data.

Team Name
XMU ATR

Reference person
Lingxiang Zheng

Affiliation

Xiamen University (CN)

Description

The XMU_PDR system is a multi-sensor fusion system for indoor positioning. It fuses the information provided by inertial measurement unit, magnetometer, barometer, WIFI, GNSS and light sensor and the maps. It calculates the PDR trajectory using the data of accelerometer and gyroscope firstly. By using all the sensors information, the ubiquitous context awareness positions were found out to eliminate the accumulated drift error. The magnetic fingerprint was collected using the sensor data in the training set. The dynamic time warping algorithm is used to match the trajectory with the trajectories in the training set. At the same time, the trajectory is corrected according to the map information, and the optimal estimation under the map constraint is realized according to the constraint of walls and special positions such as the staircase and doorway. The barometer data and WiFi data in the training set are also used to predict the changes of floor.

Team Name
IOT2US

Reference person
Bang Wu

Affiliation

Shenzhen University (CN), Queen Mary University of London and University College London (UK)

Description

1. WiFi Fingerprint positioning: for non-machine-learning algorithms, in the offline phase a RSSI ranking-based hierarchical radio map is constructed, with help of AP selection algorithms; in the online phase, RSSI ranking-based location fingerprinting method that uses Kendall Tau Correlation Coefficient (KTCC) and Convolutional Neural Networks (CNN) are used to determine the user’s location, and magnetic field-based path matching is used.
For machine/deep learning algorithms, we build the training set by using all APs’ RSSI information at known points; training our models (e.g SVM, XGBoost, KNN, CNN); predicting the positions of unknown points.
2. Pedestrian Dead Reckoning (PDR): step counter; pose estimator updated by using the quaternion method, besides, AHRS contains the information of the quaternion, which can be used to fuse with estimated heading.
3. Mobility Detection: we build the training set by collecting the data from provided training and validate sets; Train our models (e.g SVM, XGBoost, DNN, CNN); predict the mobility modes of the user’s state.
4. Floor Level Detection: similar to mobility detection.
5. Indoor/Outdoor Detection: we use GNSS to discriminate indoor scenarios and outdoor scenarios using machine learning algorithms.
6. Data Fusion: extended Kalman Filter or Unascend Kalman Filter or Particle Filter.

Team Name
WiMaP

Reference person
Qu Wang

Affiliation

Beijing University of Posts and Telecommunications; Institute of Computing Technology, Chinese Academy of Sciences (CN)

Description

We use a neural network-based indoor positioning based on pedestrian dead reckoning and particle filter to account for the influence on positioning of the indoor spatial map. In the training process, the smartphone built-in sensors are used as data sources to provide fine-grained position observations. The sparse Wi-Fi signal are trained for regressing a position which label is provide by PDR mothed. In the test process, Wi-Fi signal are used as observations in the particle filter.

Track 4 - Foot-mounted IMU (off-site)

Team Name

WHUGNSS

Team members
Jian Kuang, Tao Liu, Yan Wang

Affiliation
Wuhan University (CN)

Description

The classic zero-velocity update algorithm (ZUPT) based foot-mounted pedestrian dead reckoning consists of a strap-down inertial navigation algorithm, a stance phase detection algorithm, and an error state Kalman filter. However, the classic ZUPT-based Foot-PDR cannot be overcome the influence of the complex motion of the pedestrian. Several schemes are designed to improve navigation performance. 1) An improved adaptive threshold algorithm to detect the stance-phase in each gait cycle. 2) The zero angular rate update (ZARU) algorithm, the improved heuristic drift elimination (iHDE), and the straight-line constraint algorithm are used to constraint the heading error drift. 3) A motion detection algorithm is used to distinguish ground, escalator, and elevator, and a constant speed constraint is used to update the velocity vector when the pedestrian takes the escalator and elevator. 4) The calibrated magnetometer observations are used to detect whether the user returns to the same area. 5) The loosely integrated model is used to combine Foot-PDR and GNSS signals, and an adaptively robust algorithm is used to improve the performance of the Kalman filter. 6) The optimal inertial sensor parameters (i.e., the bias instability of gyroscopes and accelerometers, the angular random walk, and the velocity random walk) are determined through the provided long-term static data.

Team Name
BHSNIP

Team members
Ming Xia, Dayu Yan, Yuhang Li, Yitong Dong and Haitao Jiang

Affiliation
Beihang University (CN)

Description

Researchers usually suppress the divergence of positioning error within IEZ (Inertial Navigation System - Extended Kalman Filter - Zero Velocity Update, INS-EKF-ZUPT) framework. However, due to the poor observability of heading errors to ZUPT and the instability of vertical inertial channels, further corrections of the estimated trajectories under the IEZ framework are still needed to obtain higher positioning accuracy. First, the improved Step Height Equidistant (SHE) method was exploited to estimate the altitude, which was based on vertical motion modes. Then, the Strapdown Inertial Navigation System (SINS) and GPS was integrated according to loose combination. Furthermore, the improved heuristic drift elimination was also used to constraint the heading drift. Finally, the adaptive method was utilized to adjust the parameters of EKF and the threshold of ZUPT.

Team Name
Free-Walking

Team members
Haiyong Luo, Qu Wang

Affiliation
Beijing University of Posts and Telecommunications; Institute of Computing Technology, Chinese Academy of Sciences (CN)

Description

A foot-mounted pedestrian inertial navigation system that accurately tracks pedestrian position by using inertial measurement units (IMUs) embedded in device when pedestrian walking normally. However, the positioning accuracy is decreased under complex movements. In this work, we divide the commonly used indoor walking motion type into eight modes construct motion mode classifier based on stacked denoising autoencoder and temporal convolutional network with attention to recognize these pedestrian motion modes. Base on the walking mode classification, we optimize the threshold or parameters of strapdown inertial navigation and zero-velocity detection, and Kalman filter for each walking mode. The proposed method can more effectively distinguish the pedestrian’s walking mode, accurately detect the stationary phase and estimate the device attitude under various pedestrian movements. The results of multi-floor and mixed walking modes experiment show that the positioning errors of the proposed method are less than 2%.

Team Name
AIR

Reference person
Wenchao Zhang

Affiliation
Aerospace Information Research Institute, Chinese Academy of Sciences (CN)

Description

Still Phase Detection includes two components: the GLRT detector algorithm used under the condition of the slow and normal pedestrian gait speed, and the HMM detector algorithm used under the condition of the dynamic and fast pedestrian gait speed. After that, using the improved HDE and HUPT method to estimate current position errors, ZUPT is used to estimate the velocity error, while Earth Magnetic Yaw based on QSF method is used to estimate the heading error.

A gait or a walk cycle consists of two phases: the swing and stance phase. In the swing phase, the foot is not in contact with the ground. In contrast, the foot contacts the ground in the stance phase  GLRT algorithm has obvious advantages for zero speed detection of stable pedestrian gait velocity, while HMM algorithm has a good effect for zero speed detection of dynamic and fast pedestrian gait speed  Thus, the two methods are combined to achieve the dynamic human stance.

Track 5 - xDR in manufacturing (off-site)

Team Name
Kawaguchi Lab

Reference person
Takuto Yoshida

Affiliation
Nagoya university (JP)

Description

A robust xDR for a target to track based on a velocity estimation method using a neural network. A trajectory estimation by this method consists of the following three steps: 1) an end-to-end speed estimation using a neural network based on LSTM. 2) A heading estimation by integrating angular velocity on the z-axis which is a projection of angular velocity to gravity. 3) Correcting the trajectory calculated from speed and heading using BLE beacon signal, reference point, and map. Our challenge is how to extend the velocity estimation method originally intended for PDR to VDR.

Team Name
yonayona

Reference person
Yoshitomo Yonamoto

Affiliation
Keio university (JP)

Description

My main method is to emphasis on absolute positioning technology. In a certain of time, I use more than 3 detected beacons, try various approaches such as average weighted and trilateration. and choose best estimation.
I also tackle to decrease negative check points by implementing prevent-intersect-wall algorithm.

Track 6 - On-vehicle smartphone (off-site)

Team Name
SZU Mellivora Capensis

Reference person
Xu Liu

Affiliation
Shenzhen University (CN)

Description

Our system proposes a novel method (named SAM-AI Location System) that integrated an improved deep recurrent neural network-based method with a traditional 3D inertial dead-reckoning method to achieve high-precision positioning using the sensor data obtained from smart phone.
The system transforms the coordinate system according to the attitude data, and integrates GNSS information when it detects GNSS signals. The system uses the dead reckoning inertial method based on IMU to estimate the three-dimensional position, speed and direction of the vehicle. At the same time, we use the deep recurrent neural networks to learn the data of accelerometer, gyroscope and magnetometer to train model and predict the highly accurate vehicle track. The inertial-based track obtained by the integrated method through federated filter. Moreover, this system adds GPS information where it detects a GPS signal. GPS signal provide starting information for our inertial-based algorithms. It is worth mentioning that we consider the transformation of coordinate system and the installation angle of smart phone in the process of deducing trajectory.The positioning results of the competition show that the SAM-AI Location System proposed by us can solve the positioning requirements in the complex indoor and outdoor environment.

Team Name
YAI

Reference person
Chia An

Affiliation
Department of Electrical Engineering, Yuan Ze University and National Ilan University (TW)

Description

The systems is based on dead reckoning. We use Kalman filter to solve the problem of straight driving, and use Stanley control algorithm to solve the problem of deviation when turning. The data provided by the database uses ACCE, GYRO, AHRS, GNSS.
We use the road section with GNSS signal to train parameters.

Team Name
WHU-Autonavi

Reference person
Jian Kuang

Affiliation

Wuhan university;  AutoNavi Software Co., Ltd. (CN)

Description

 A smartphone has many kinds of sensors but their performance is poor, and the smartphone bracket is easy to shake, furthermore, the GNSS signal is interrupted frequently. Therefore, we add the following schemes for the traditional data processing strategy of the vehicle GNSS/INS integrated navigation system. 1) Adjust and estimate the performance parameter of smartphone built-in sensors; 2) Estimate the misalignment angles between smartphone and vehicle; 3) Estimate non-holonomic constraint (NHC) lever arm; 4) Detect the static period of the vehicle and implement ZUPT and ZARU;5) Using magnetic field to restrain heading drift;6) Establish machine learning model to estimate vehicle forward speed. Through the above processing, our algorithm can give continuous and effective positioning results of the scene.

Track 7 - Channel impulse response (off-site)

Team Name
YAI

Reference person
NienTing Lee

Affiliation
Department of Electrical Engineering, Yuan Ze University and National Ilan University (TW)

Description

We use neural network to train the model for each receiver, before training model, we process the square and root of the real and imaginary parts of the CIR signal, and then feed the processed 366 signals to neural network training.
For labeling, we take the integer digits, and treat each point mark as a category, and finally divide a total of 285 categories.

Competing Teams

14 December 2020 - Zoom workshop

Track 3 - Smartphone (off-site)

Team Name
Next-Newbie Reckoners

Reference person

Seanglidet Yean

Affiliation

Singtel Cognitive and Artificial Intelligence Lab for Enterprises at NTU (SCALE@NTU); Nanyang Technological University (SG)

Description

We have adopted grid-based approach to train the Random Forest models with an end-to-end pipleline in order to predict the building, floor and location grid using the WiFi Received Signal Strength. In addition, we are working on the Deep Neural Network models as well as feature selection methods to further improve the accuracy.
In this competition, our main objective is to incorporate multi-source of information such as Bluetooth, Smartphone sensor-fusion techniques (e.g. pedestrian dead reckoning) to enhance our method.

Team Name

WHU-five

Reference person

Li Wei

Affiliation
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (LIESMARS), WuHan University (CN)

Description

Training: According to the training data set, establish a geomagnetic fingerprint library, a WIFI fingerprint library, and a lighting library.
Test: Estimate the initial position through WIFI fingerprint positioning, lights, geomagnetism, etc., and then merge the results of PDR, geomagnetism, and wifi positioning to get the final position.
The floor can be judged roughly based on the barometer. When there is a weak change in air pressure between floors, the floor information can be detected by fusing other signals (wifi, geomagnetism, lights, etc.).

Team Name

YAI

Reference person
ChihChieh Yu

Affiliation
Department of Electrical Engineering, Yuan Ze University and National Ilan University (TW)

Description

Our fingerprinting is based on Wi-Fi signal strength (RSS) values expressed in dBm. In this competition, we propose a fuzzy-based pre-processing method so that the RSS entries in the fingerprinting database can be converted into the corresponding defuzzification values. The fuzzy-based pre-processing utilizes five membership functions on the received RSS signals and the parameters associated with these membership functions are trained by the training dataset. Finally, we obtain a non-linear mapping function, which mapping the RSS values to the real values between the range of 0 to 4 as shown in Fig. 1. More specifically, we adopt air pressure value for floor judgment, acceleration value for step detection, and Wi-Fi signal for fingerprinting table positioning.

Team Name
Indora

Reference person
Miroslav Opiela

Affiliation
Institute of Computer Science, P.J. Šafárik University, Faculty of Science, Jesenná 5, 041 54 Košice (SK)

Description

Pedestrian dead-reckoning, map model and bayesian filtering are essential components of the proposed localization system. The main research focus is on a low-dimensional grid-based bayesian filtering, as a less elaborated alternative to Kalman and Particle filters widely used for the positioning. A semi-automatically generated map model helps to reduce the localization error introduced by noisy sensor measurements and inaccurate system configuration, e.g., a step length estimation. The floor transition detection based on barometer measurements allows the system to estimate the user position within a single floor. Step length estimation is incorporated in the particle filter. The light sensor improves the ability to distinguish between two similar corridors and detect indoor-outdoor transitions.

Team Name
UMinho

Reference person
Ivo Silva

Affiliation

University of Minho (PT)

Description

The solution is based on a Wi-Fi radio map built with the signals available in the training dataset. It uses sensor fusion to merge the results obtained from the Wi-Fi fingerprinting with data from other sensors, including accelerometers records to detect the person steps, the mobile phone 3D orientation records for heading estimation and pressure sensors for floor transition detection.

Team Name
imec-WAVES/Ghent University

Reference person
Cedric De Cock

Affiliation

imec-WAVES/Ghent University (BE)

Description

Our system consists of 4 parts. The first part uses intertial sensordata to detect steps, estimate step length and heading estimation and to recognize walking modes. The second part uses the PDR algorithm to interpolate between the sparse ground truth positions. The interpolated positions are used to create a radiomap. The third part uses the barometer and RSS measurements from the training data to detect floor transitions, to estimate during which steps the user is using the stairs and to calculate the most likely sequence of floors visited by the user. In the final part, a backtracking particle filter (BPF) is used to estimate the trajectory. The step lengths and headings are used to propagate the particles. The positions of walls are used to remove particles that were propagated through a wall. The weighted centroid of the k best RSS matches is used to weigh each particle, based on its distance to the matched position. The locations of staircases are used to decrease the weight of particles outside of staircases when the barometer detects a height change. The floor transitions are used to load the next floor map and radiomap.

Team Name
TJU

Team members
Liqiang Zhang, Boxuan Chen, Hu Li, Yazhen liao, Qingyuan Gong and Yu Liu

Affiliation

School of Microelectronics, Tianjin University (CN)

Description

A smartphone integrates rich sensors, including inertial sensor, magnetic sensor, light sensor, sound sensor and proximity sensor, etc. Therefore, it provides a possibility to localize pedestrians, which is valuable and challenging. In this competition, step-length and heading system (SHS) is used as our fundamental system. Acceleration information from inertial sensor is used to detect steps. Then the step length can be computed. To our best knowledge, step-length computing is accurate when a pedestrian moves forward on a level ground. However, the estimated step length becomes inaccurate when a pedestrian moves on a slope, such as walking upstairs or downstairs. To address the problem, we propose adaptive step-length estimation method aided by learning-based motion classifier. Since the performance of the SHS degrades with time and walking distance, reference positions should be introduced to modify the positioning error of the system. Therefore, we establish magnetic field and Wi-Fi-based fingerprint databases with the training and validation data. Then a neural network (NN) model is trained to predict the pre-given ground truth positions using geomagnetic information and Wi-Fi received signal strength. Finally, Kalman filter is leveraged to fuse the positioning results from the NN model and the SHS.

Team Name
MAD PDR (Magnetic Assisted Deep PDR)

Reference person
Leonid Antsfeld

Affiliation

Naver Labs Europe (FR)

Description

We have used a stack of complementary localization components that were intelligently fused together to provide a final actor’s itinerary.
The main components are: floor detection, user activity recognition, speed/step length estimation, PDR, WiFi based localization and, finally, a novel component of magnetic field based localization.
Floor detection was done using classic classification methods based on WiFi and barometer data. We have applied spectral analysis on the accelerometer data in order to identify user's activity (walking, standing, going up or down the stairs). Next, we have used peaks detection of accelerometer data in order to detect steps. We extract acceleration features in order to learn the step length / speed in a given sliding window. Together with the orientation sensor it allowed us to build a PDR which was a first order approximation of the itinerary.
The PDR is known for being accurate locally but drifting over time. In order to compensate for this drift, we have used global localization techniques based on WiFi VAE that was presented at IPIN’19 and recently developed LSTM based magnetic field localization algorithm.
Finally, the itinerary was adjusted using a map that was extracted from the training data.

Team Name
XMU ATR

Reference person
Lingxiang Zheng

Affiliation

Xiamen University (CN)

Description

The XMU_PDR system is a multi-sensor fusion system for indoor positioning. It fuses the information provided by inertial measurement unit, magnetometer, barometer, WIFI, GNSS and light sensor and the maps. It calculates the PDR trajectory using the data of accelerometer and gyroscope firstly. By using all the sensors information, the ubiquitous context awareness positions were found out to eliminate the accumulated drift error. The magnetic fingerprint was collected using the sensor data in the training set. The dynamic time warping algorithm is used to match the trajectory with the trajectories in the training set. At the same time, the trajectory is corrected according to the map information, and the optimal estimation under the map constraint is realized according to the constraint of walls and special positions such as the staircase and doorway. The barometer data and WiFi data in the training set are also used to predict the changes of floor.

Team Name
IOT2US

Reference person
Bang Wu

Affiliation

Shenzhen University (CN), Queen Mary University of London and University College London (UK)

Description

1. WiFi Fingerprint positioning: for non-machine-learning algorithms, in the offline phase a RSSI ranking-based hierarchical radio map is constructed, with help of AP selection algorithms; in the online phase, RSSI ranking-based location fingerprinting method that uses Kendall Tau Correlation Coefficient (KTCC) and Convolutional Neural Networks (CNN) are used to determine the user’s location, and magnetic field-based path matching is used.
For machine/deep learning algorithms, we build the training set by using all APs’ RSSI information at known points; training our models (e.g SVM, XGBoost, KNN, CNN); predicting the positions of unknown points.
2. Pedestrian Dead Reckoning (PDR): step counter; pose estimator updated by using the quaternion method, besides, AHRS contains the information of the quaternion, which can be used to fuse with estimated heading.
3. Mobility Detection: we build the training set by collecting the data from provided training and validate sets; Train our models (e.g SVM, XGBoost, DNN, CNN); predict the mobility modes of the user’s state.
4. Floor Level Detection: similar to mobility detection.
5. Indoor/Outdoor Detection: we use GNSS to discriminate indoor scenarios and outdoor scenarios using machine learning algorithms.
6. Data Fusion: extended Kalman Filter or Unascend Kalman Filter or Particle Filter.

Team Name
WiMaP

Reference person
Qu Wang

Affiliation

Beijing University of Posts and Telecommunications; Institute of Computing Technology, Chinese Academy of Sciences (CN)

Description

We use a neural network-based indoor positioning based on pedestrian dead reckoning and particle filter to account for the influence on positioning of the indoor spatial map. In the training process, the smartphone built-in sensors are used as data sources to provide fine-grained position observations. The sparse Wi-Fi signal are trained for regressing a position which label is provide by PDR mothed. In the test process, Wi-Fi signal are used as observations in the particle filter.

Track 4 - Foot-mounted IMU (off-site)

Team Name
WHUGNSS

Team members
Jian Kuang, Tao Liu, Yan Wang

Affiliation
Wuhan University (CN)

Description

The classic zero-velocity update algorithm (ZUPT) based foot-mounted pedestrian dead reckoning consists of a strap-down inertial navigation algorithm, a stance phase detection algorithm, and an error state Kalman filter. However, the classic ZUPT-based Foot-PDR cannot be overcome the influence of the complex motion of the pedestrian. Several schemes are designed to improve navigation performance. 1) An improved adaptive threshold algorithm to detect the stance-phase in each gait cycle. 2) The zero angular rate update (ZARU) algorithm, the improved heuristic drift elimination (iHDE), and the straight-line constraint algorithm are used to constraint the heading error drift. 3) A motion detection algorithm is used to distinguish ground, escalator, and elevator, and a constant speed constraint is used to update the velocity vector when the pedestrian takes the escalator and elevator. 4) The calibrated magnetometer observations are used to detect whether the user returns to the same area. 5) The loosely integrated model is used to combine Foot-PDR and GNSS signals, and an adaptively robust algorithm is used to improve the performance of the Kalman filter. 6) The optimal inertial sensor parameters (i.e., the bias instability of gyroscopes and accelerometers, the angular random walk, and the velocity random walk) are determined through the provided long-term static data.

Team Name
BHSNIP

Team members
Ming Xia, Dayu Yan, Yuhang Li, Yitong Dong and Haitao Jiang

Affiliation
Beihang University (CN)

Description

Researchers usually suppress the divergence of positioning error within IEZ (Inertial Navigation System - Extended Kalman Filter - Zero Velocity Update, INS-EKF-ZUPT) framework. However, due to the poor observability of heading errors to ZUPT and the instability of vertical inertial channels, further corrections of the estimated trajectories under the IEZ framework are still needed to obtain higher positioning accuracy. First, the improved Step Height Equidistant (SHE) method was exploited to estimate the altitude, which was based on vertical motion modes. Then, the Strapdown Inertial Navigation System (SINS) and GPS was integrated according to loose combination. Furthermore, the improved heuristic drift elimination was also used to constraint the heading drift. Finally, the adaptive method was utilized to adjust the parameters of EKF and the threshold of ZUPT.

Team Name
Free-Walking

Team members
Haiyong Luo, Qu Wang

Affiliation
Beijing University of Posts and Telecommunications; Institute of Computing Technology, Chinese Academy of Sciences (CN)

Description

A foot-mounted pedestrian inertial navigation system that accurately tracks pedestrian position by using inertial measurement units (IMUs) embedded in device when pedestrian walking normally. However, the positioning accuracy is decreased under complex movements. In this work, we divide the commonly used indoor walking motion type into eight modes construct motion mode classifier based on stacked denoising autoencoder and temporal convolutional network with attention to recognize these pedestrian motion modes. Base on the walking mode classification, we optimize the threshold or parameters of strapdown inertial navigation and zero-velocity detection, and Kalman filter for each walking mode. The proposed method can more effectively distinguish the pedestrian’s walking mode, accurately detect the stationary phase and estimate the device attitude under various pedestrian movements. The results of multi-floor and mixed walking modes experiment show that the positioning errors of the proposed method are less than 2%.

Team Name
AIR

Reference person
Wenchao Zhang

Affiliation
Aerospace Information Research Institute, Chinese Academy of Sciences (CN)

Description

Still Phase Detection includes two components: the GLRT detector algorithm used under the condition of the slow and normal pedestrian gait speed, and the HMM detector algorithm used under the condition of the dynamic and fast pedestrian gait speed. After that, using the improved HDE and HUPT method to estimate current position errors, ZUPT is used to estimate the velocity error, while Earth Magnetic Yaw based on QSF method is used to estimate the heading error.

A gait or a walk cycle consists of two phases: the swing and stance phase. In the swing phase, the foot is not in contact with the ground. In contrast, the foot contacts the ground in the stance phase  GLRT algorithm has obvious advantages for zero speed detection of stable pedestrian gait velocity, while HMM algorithm has a good effect for zero speed detection of dynamic and fast pedestrian gait speed  Thus, the two methods are combined to achieve the dynamic human stance.

Track 5 - xDR in manufacturing (off-site)

Team Name
Kawaguchi Lab

Reference person
Takuto Yoshida

Affiliation
Nagoya university (JP)

Description

A robust xDR for a target to track based on a velocity estimation method using a neural network. A trajectory estimation by this method consists of the following three steps: 1) an end-to-end speed estimation using a neural network based on LSTM. 2) A heading estimation by integrating angular velocity on the z-axis which is a projection of angular velocity to gravity. 3) Correcting the trajectory calculated from speed and heading using BLE beacon signal, reference point, and map. Our challenge is how to extend the velocity estimation method originally intended for PDR to VDR.

Team Name
yonayona

Reference person
Yoshitomo Yonamoto

Affiliation
Keio university (JP)

Description

My main method is to emphasis on absolute positioning technology. In a certain of time, I use more than 3 detected beacons, try various approaches such as average weighted and trilateration. and choose best estimation.
I also tackle to decrease negative check points by implementing prevent-intersect-wall algorithm.

Track 6 - On-vehicle smartphone (off-site)

Team Name
SZU Mellivora Capensis

Reference person
Xu Liu

Affiliation
Shenzhen University (CN)

Description

Our system proposes a novel method (named SAM-AI Location System) that integrated an improved deep recurrent neural network-based method with a traditional 3D inertial dead-reckoning method to achieve high-precision positioning using the sensor data obtained from smart phone.
The system transforms the coordinate system according to the attitude data, and integrates GNSS information when it detects GNSS signals. The system uses the dead reckoning inertial method based on IMU to estimate the three-dimensional position, speed and direction of the vehicle. At the same time, we use the deep recurrent neural networks to learn the data of accelerometer, gyroscope and magnetometer to train model and predict the highly accurate vehicle track. The inertial-based track obtained by the integrated method through federated filter. Moreover, this system adds GPS information where it detects a GPS signal. GPS signal provide starting information for our inertial-based algorithms. It is worth mentioning that we consider the transformation of coordinate system and the installation angle of smart phone in the process of deducing trajectory.The positioning results of the competition show that the SAM-AI Location System proposed by us can solve the positioning requirements in the complex indoor and outdoor environment.

Team Name
YAI

Reference person
Chia An

Affiliation
Department of Electrical Engineering, Yuan Ze University and National Ilan University (TW)

Description

The systems is based on dead reckoning. We use Kalman filter to solve the problem of straight driving, and use Stanley control algorithm to solve the problem of deviation when turning. The data provided by the database uses ACCE, GYRO, AHRS, GNSS.
We use the road section with GNSS signal to train parameters.

Team Name
WHU-Autonavi

Reference person
Jian Kuang

Affiliation

Wuhan university;  AutoNavi Software Co., Ltd. (CN)

Description

 A smartphone has many kinds of sensors but their performance is poor, and the smartphone bracket is easy to shake, furthermore, the GNSS signal is interrupted frequently. Therefore, we add the following schemes for the traditional data processing strategy of the vehicle GNSS/INS integrated navigation system. 1) Adjust and estimate the performance parameter of smartphone built-in sensors; 2) Estimate the misalignment angles between smartphone and vehicle; 3) Estimate non-holonomic constraint (NHC) lever arm; 4) Detect the static period of the vehicle and implement ZUPT and ZARU;5) Using magnetic field to restrain heading drift;6) Establish machine learning model to estimate vehicle forward speed. Through the above processing, our algorithm can give continuous and effective positioning results of the scene.

Track 7 - Channel impulse response (off-site)

Team Name
YAI

Reference person
NienTing Lee

Affiliation
Department of Electrical Engineering, Yuan Ze University and National Ilan University (TW)

Description

We use neural network to train the model for each receiver, before training model, we process the square and root of the real and imaginary parts of the CIR signal, and then feed the processed 366 signals to neural network training.
For labeling, we take the integer digits, and treat each point mark as a category, and finally divide a total of 285 categories.