The EvAAL Framework
Competing Teams 2023
Offsite: 11-15 Sep, EvaalAPI web service
Track 1 - Smartphone (onsite) Team Name
THWS
Team members Markus Ebner, Steffen Kastner, Markus Bullmann, Toni Fetzer, Frank Deinzer
Affiliation Technische Hochschule Würzburg-Schweinfurt
Description Based on a condensation particle filter, the localization system is highly modular with respect to the sensors it is able to use.
Models exist for Bluetooth low energy beacons, Wi-Fi, and others. Combined with a step and turn detection, it can also handle scenarios of total sensor outages. All calculations are performed in real-time on a commercial smartphone even when using a high number of samples for approximation.  Team Name
Inha & Papaya
Team members Wonik Choi, Chungheon Yi , Taeyeab Kim, Young-jun Jeon, Minjoon Jeon, Kiryang Kwon
Affiliation Inha University, Papaya Co., LTD.
Description PapayaIPS is an artificial intelligence-based positioning solution, which employs a specially designed comparative neural network to estimate locations. We develop our own comparative neural network and call this neural network PNN (Positioning Neural Network). PapayaIPS is basically aiming to utilize all signals that can be received by a smartphone such as Wi-Fi, BLE, barometer, and accelerometer, etc., to provide a comprehensive positioning solution.
PapayaIPS is based on signal tensor images that are transformed and visualized from those multi-dimensional signal values received by a smartphone at a certain point. Then, PNN learns to transform the similarity of two signal tensors into the Euclidean distance indoors by training with these signal tensor images. Team Name
KNU WCSL
Team members Sumin Joo, Hyeonseon An, Younghun Ha, Jeongsik Choi
Affiliation Kyungpook National University
Description Track 3 - Smartphone (offsite-online) Team Name
IOT2US Team members Yonglei Fan, Qiqi Shu, Zhao Huang, Guangyuan Zhang, Meng Xu, Xijie Xu, Guangxia Yu, Zhao liang Luan, Stefan Poslad
Affiliation Queen Mary, University of London, Peking University
Description it is real time indoor positioning method, we will separate all data into small pieces and use these pieces of data to do the positioning. For the first positioning, we may use WiFi, or Bluetooth data. Kalman filter will be used to minimize the error and the cumulative error. But there must be some cumulative error which is hard to deal with. We should think a better way to fix it. Pedestrian dead reckoning (PDR) has become a research hotspot since it does not require a positioning infrastructure. An integral equation is used in PDR positioning; thus, errors accumulate during long-term operation. To eliminate the accumulated errors in PDR localisation, we proposes a PDR localisation system applied to complex scenarios with multiple buildings and large areas. The system is based on the pedestrian movement behavior recognition algorithm proposed. Team Name
IA3 Team members Alex Martinez-Martinez, Javier Gabiña Rueda
Affiliation Universitat Jaume I
Description The system aims to achieve accurate indoor positioning through the integration of various technologies such as Pedestrian Dead Reckoning (PDR), WiFi fingerprinting, multi-sensor fusion, activity recognition, and map information. Core components include multi-sensor fusion, WiFi fingerprinting, and PDR prediction. These components work together to predict the user's position using sensors and WiFi data. Activity recognition and map information complement the predictions. Machine learning models such as LSTMs, RNNs, and CNNs are employed. The system enhances accuracy by identifying floor changes and predicting turns. It utilizes map information to guide users and ensure they stay within building boundaries. The goal is to provide highly accurate and reliable indoor positioning. Team Name
QAQ Team members Shiyu Zheng, Yuwen Guo, Ziheng Zhou, Qi Zhang, Shanshan Zhang, Lingxiang Zheng, Ao Peng
Affiliation Xiamen University
Description Our multi-source indoor positioning system integrates data from an Inertial Measurement Unit (IMU), Wi-Fi, magnetometer, barometer, and indoor maps to achieve precise indoor localization. The system utilizes the Inverted Pendulum Model-based Pedestrian Dead Reckoning (PDR) technique to estimate the user's position. In addition to PDR, the system incorporates Wi-Fi and Bluetooth data to create fingerprint maps. Fingerprint matching is employed to determine the user's current position, while changes in barometric pressure aid in floor level detection. The data from each component is fused using a factor graph-based approach, specifically based on gtsam, to achieve accurate and reliable positioning results. Team Name
Inha & Papaya Team members Wonik Choi, Chungheon Yi , Taeyeab Kim, Young-jun Jeon, Minjoon Jeon, Kiryang Kwon
Affiliation Inha University, Papaya Co., LTD.
Description PapayaIPS is an artificial intelligence-based positioning solution, which employs a specially designed comparative neural network to estimate locations. We develop our own comparative neural network and call this neural network PNN(Positioning Neural Network). PapayaIPS is basically aiming to utilize all signals that can be received by a smartphone such as Wi-Fi, BLE, barometer, and accelerometer, etc., to provide a comprehensive positioning solution. PapayaIPS is based on signal tensor images that are transformed and visualized from those multi-dimensional signal values received by a smartphone at a certain point. Then, PNN learns to transform the similarity of two signal tensors into the Euclidean distance indoors by training with these signal tensor images. Team Name KNU WCSL Team members Sumin Joo, Hyeonseon An, Younghun Ha, Jeongsik Choi
Affiliation Kyungpook National University
Description We consider a well-known positioning approach which is based on Wi-Fi ranging and PDR method. Such an approach requires to secure the location database of Wi-Fi access points (APs) and investigate the signal propagation characteristics under each site for precise ranging. To minimize human intervention to conduct such a site survey task, we apply an automatic site survey method that estimates the locations of APs and learn signal propagation characteristics simultaneously using measurement data provided from multiple mobile users. Team Name Team members
Mengyuan Tang
Affiliation Shenzhen University, College of Civil and Transportation Engineering
Description Our system uses deep recurrent neural networks to learn the sensor data from smartphones. Based on sensor data and BLE beacons information, we train to recognize the basic behaviors of pedestrians inside buildings (detecting behaviors like going upstairs, in a lift, turning, etc.), and integrate the attitude results of pedestrians dead reckoning (PDR) methods to achieve trajectory positioning of pedestrians in multi-floor buildings. Nowadays, Indoor space has been an important space for human activities and people spend more than 80% of their time in the indoor environment. However, as a common scenario in urban indoor spaces, multi-floor buildings face issues such as missing or attenuated GNSS signals. Therefore, it is one of the current research hotspots to realize the robust low-cost navigation and positioning in complex indoor environments. Our system uses the PDR methods based on the Inertial Measurement Unit (IMU) to estimate the three-dimensional position. Team Name
BUPTer Team members Chaoyi Xu, Fan Yin
Affiliation Beijing University of Post and Telecommunication, China (BUPT)
Description Our project aims to utilize an Inertial Measurement Unit (IMU) as the principal navigation sensor, augmented by additional sensors, to realize precise positioning. Notably, the inherent integration process of Pedestrian Dead Reckoning (PDR) might cause cumulative errors, especially when mobile-phone grade IMUs, which are less accurate and more prone to individual's walking styles, are employed. Hence, our project proposes to incorporate WiFi fingerprint data to assist the inertial navigation system, leveraging a fuzzy matching technique to mitigate the accumulation of errors. We also plan to implement a filtering and fusion algorithm to reduce disturbances from electromagnetic interference or sensor errors. Further, we use a subset of actual location data to intermittently calibrate for path deviations. It may also involve the preliminary conversion of longitude and latitude coordinates into a planar coordinate system. Team Name
THWS
Team members Markus Ebner, Steffen Kastner, Markus Bullmann, Toni Fetzer, Frank Deinzer
Affiliation Technische Hochschule Würzburg-Schweinfurt
Description Based on a condensation particle filter, the localization system is highly modular with respect to the sensors it is able to use. Models exist for Bluetooth low energy beacons, Wi-Fi, and others. Combined with a step and turn detection, it can also handle scenarios of total sensor outages. All calculations are performed in real-time on a commercial smartphone even when using a high number of samples for approximation.  Team Name
imec-Waves
Team members Cedric De Cock, David Plets
Affiliation Ghent University/imec-Waves
Description Our system consists of six modules, which are briefly described in this document. The first module is an interface for the Evaal server, which starts the Evaal trial and requests the next stream of smartphone data in blocks of 0.5 s, and uploads the estimated position. The second modules is a Pedestrian Dead reckoning (PDR) algorithm, which detects steps from the smartphone’s IMU data, and estimates the step length and heading. The third module is a 3D graph of the environment, in which the nodes are accessible positions, and the edges are physically valid (for a human) ways of traveling between the nodes. The fourth module performs BLE fingerprinting, matching new RSS measurements using KNN. The fifth modules is a floor (transition) detection algorithm, which (f)uses barometer, accelerometer, GPS, RSS data. The sixth module is a tracking algorithm which incorporates the graph, PDR, fingerprinting, and floor (transition) detection to estimate the user position. Track 4 - Foot-mounted IMU (offsite-online)Team Name
ININ624
Team members Zhidong Meng, Zhe Li, Ping Zhang
Affiliation School of Automation, Beijing Institute of Technology
Description The Calibration of sensors is carried out leveraging the given data including constant bias estimation, temperature compensation, and random characteristics for the gyroscope. The calibration of magnetometer is realized by ellipsoid fitting method. The initial velocity is set to 0, and the initial position is obtained by GNSS solution. The initial attitude is determined by TRIAD method. The error state sequential Kalman filter is used to estimate the errors in attitude, velocity, position, accelerometer drifts, gyroscope drifts, and body odometer (BOR) drifts. The observation resources from pseudo Zero-velocity measurement, Barometer, Global Navigation Satellite System, and BOR. Fusing BOR is an algorithm proposed by our team that utilizes displacement estimation to suppress multiple errors in inertial navigation by breaking the integral process. To avoid failed observer to disturb measurement updates, a chi square test and expectation limitation are introduced. Team Name
CETC-CePNT
Team members Baoguo Yu, Jun Li, Xinjian Wang, Yanan Hu, Haonan Jia, Lu Huang
Affiliation The 54th Research Institute of China Electronics Technology Group Corporation
Description Multiple conditions zero-velocity detection algorithms are employed to detect zero-velocity intervals within the walking gait. Within the detected zero-velocity intervals, the principle of zero-velocity correction algorithm is utilized to construct the observed quantity of velocity error; leveraging the characteristics of stationary pedestrian foot-ground contact and only subjected to gravity acceleration and unchanged posture angles within the zero-velocity intervals to construct the observed quantity of posture angle error. Satellite signals acquired by the system are differentiated and identified, usable signals are extracted, and a Kalman filter is applied to estimate errors in posture angles, velocity, and position within the zero-velocity intervals. The obtained error state estimation results are utilized for error correction in pedestrian navigation, enhancing the accuracy of inertial pedestrian navigation. Team Name
VINF
Team members Jiale Han, Maoran Zhu, Yuanxin Wu
Affiliation Institute for Sensing and Navigation, Shanghai Jiao Tong University
Description In our system design, the core is an information fusion algorithm based on the Error State Kalman Filter with five constraints, including the Zero-velocity update (ZUPT), the Zero angular rate update (ZARU), the Improved heuristic drift elimination (iHDE), the Ellipsoid constraint, and the Constant speed. Additionally, the height variation calculated by the pressure sensor can also be used to correct the navigation state. The magnetic field will be used for loop detection, enabling the utilization of historical estimated positions to correct the current navigation state. When the user comes to an outdoor scene, the received GNSS signals will be used to correct the current navigation state.
The parameters of inertial sensors, such as the bias instability of gyroscopes and accelerometers, the angle random walk, and the velocity random walk, are determined through long-term static data. Team Name
SmartLoc
Team members Wang Han, Zhang Mingkai
Affiliation Nanyang Technological University
Description Indoor localization has emerged as a crucial technology for a wide array of applications such as robotics, asset tracking, and indoor navigation. Traditional GPS-based methods face limitations when applied indoors due to signal attenuation and multipath effects. To address these challenges, an innovative indoor localization method has been developed, leveraging an error-state Kalman filter that integrates data from Inertial Measurement Units (IMUs), Global Positioning System (GPS) information, and zero velocity updates. This comprehensive approach enhances the accuracy and reliability of indoor positioning, making it ideal for environments where GPS signals are not available or are unreliable. Team Name
X-lab
Team members Xiaodong Li, Zhi Xiong, Yan Cui, Yinshou Sun, Yunong Qian
Affiliation Nanjing University of Aeronautics and Astronautics, College of Automation
Description This team belongs to the Navigation Research Center (NRC) in Nanjing University of Aeronautics and Astronautics, College of Automation. Our team has long been engaged in research related to indoor positioning and pedestrian navigation technologies.
Track 5 - Smartphone (offsite-online)Team Name
BUPTer
Team members Fan Yin, Chaoyi Xu
Affiliation Beijing University of Post and Telecommunication, China (BUPT)
Description Our project aims to achieve precise positioning by utilizing an Inertial Measurement Unit (IMU) as the primary navigation sensor, along with additional sensors such as Bluetooth Low Energy (BLE) and Received Signal Strength Indicator (RSSI). However, we acknowledge that the inherent integration process of Pedestrian Dead Reckoning (PDR) using mobile-phone grade IMUs, which are less accurate and more prone to individual walking styles, can lead to cumulative errors. Therefore, we propose to incorporate BLE and RSSI data to assist the inertial navigation system, leveraging a fuzzy matching technique to mitigate the accumulation of errors. To reduce disturbances from electromagnetic interference or sensor errors, we plan to implement a filtering and fusion algorithm using only the available accelerometer and gyroscope data. Team Name
SmartLoc
Team members Wang Han
Affiliation Nanyang Technological University
Description Basically, in this track the main challenge is to use the inertial information along with the beacon and map. The system is designed based on ROS and C. We leverage an extended Kalman filter and particle filer to achieve the real-time localization. Then, on the second test, where bluetooth is provided, we form a graph system. a node if placed every 1 second, with the bluetooth measurement result. we minimize the displacement error and an bluetooth penalty if the corresponding beacon is not nearby. Team Name
UCLab
Team members Kazuma Kano, Keisuke Higashiura, Kohei Yamaguchi, Koki Takigami, and Yoshiteru Nagata
Affiliation Nagoya University
Description We integrated deep-learning-based Pedestrian Dead reckoning (PDR), map-matching, multilateration, and fingerprinting on particle-filter. We analyzed heading estimation reliability and use absolute heading in reliable areas and relative heading in unreliable areas. Team Name
Kaji Lab
Team members Ryuki Toyama
Affiliation Aichi Institute of Technology
Description There are from 1 to 6 steps in the process.
1. estimate the timing of the gait from the acceleration and integrate the angular velocity to estimate the direction of motion at each time. Based on this, a two-dimensional gait trajectory is generated. 2. 2. rotate the entire walking trajectory based on the coordinates of the BLE beacon that emits strong radio waves before the half of the entire walking time. 3. perform a grid search for drift based on the final coordinates of the walking trajectory and the coordinates of the end point, and generate a walking trajectory with correction for the obtained drift. 4. rotate the entire trajectory based on the coordinates of the BLE beacon that emits strong radio waves during the entire walking time. 5. same as step 3. 6. 6. perform a simple map-matching correction. If the walking trajectory exists at a point that does not exist on the map, the trajectory is shifted to a point that does exist. Team Name
Dream Hunters
Team members Fu Han,Naoki Tanabe, Jiang Jiayin, Hiroya Yamashiro
Affiliation University of Tsukuba, Master's Programs in Intelligent and Mechanical Interaction Systems
Description This system is based on sample code aimed to estimate the indoor position of pedestrians by integrating PDR and BLE beacon reception data. We are planning to make modifications to this system before the actual competition takes place.
The estimation process utilizes IMU data (acceleration, gyro, geomagnetism) and BLE beacon reception information collected from the 9-axis IMU sensor attached to the pedestrian's AQUOS Sense 6 device (SHARP) within a commercial facility. Ultimately, accuracy is evaluated through comparison with ground truth data. Team Name
CARELab
Team members Pereira Matthieu, Huakun Liu, Yutaro Hirao, Monica Perusquia-Hernandez,Hideaki Uchiyama, and Kiyoshi Kiyokawa
Affiliation Nara Institute of Science and Technology
Description The solution proposed is based on a fingerprinting algorithm. A fingerprinting algorithm consists of making a map associating the characteristics of the data sensed with a ground truth position during a phase of training, then comparing the data sensed with the map during the testing.
Team Name
imec-Waves
Team members Cedric De Cock, David Plets
Affiliation Ghent University/imec-Waves
Description Our system consists of six modules, which are briefly described in this document. The first module is an interface for the Evaal server, which starts the Evaal trial, requests and parses the sensor data, then provides the data to the positioning system. It also uploads the estimated position sequence the Evaal server. The second module is a PDR algorithm, which detects steps from the smartphone’s IMU data, and estimates the step length and heading. The third module is a 3D graph of the environment, in which the nodes are accessible positions, and the edges are physically valid (for a human) ways of traveling between the nodes. The fourth module performs BLE fingerprinting, matching new RSS measurements using KNN. The fifth module is a floor (transition) detection algorithm, which (f)uses accelerometer and BLE data. The sixth module is a tracking algorithm which incorporates the graph, PDR, fingerprinting, and floor (transition) detection to estimate the user position. Team Name
IOT2US
Team members Yonglei Fan, Qiqi Shu, Zhao Huang, Guangyuan Zhang, Meng Xu, Xijie Xu, Guangxia Yu, Zhao liang Luan, Stefan Poslad
Affiliation Queen Mary, University of London, Peking University
Description We mainly use the acce, gyro and Mega to estimate the step length, step and direction.
it is real time indoor positioning method, we will separate all data into small pieces and use these pieces of data to do the positioning.
For the first positioning, we may use WiFi, or Bluetooth data.
Kalman filter will be used to minimize the error and the cumulative error. But there must be some cumulative error which is hard to deal with. We should think a better way to fix it.
Pedestrian dead reckoning (PDR) has become a research hotspot since it does not require a positioning infrastructure. An integral equation is used in PDR positioning; thus, errors accumulate during long-term operation. To eliminate the accumulated errors in PDR localisation, we proposes a PDR localisation system applied to complex scenarios with multiple buildings and large areas. The system is based on the pedestrian movement behavior recognition algorithm proposed.
Team Name
Royal Holloway University
Team members Xu Feng, Khuong An Nguyen, Zhiyuan Luo
Affiliation Royal Holloway University of London
Description Our system contains two main parts: the Bluetooth Low Energy (BLE) fingerprinting engine and PDR estimation engine. For BLE fingerprinting, the system uses the offline trials as training data to train a machine learning-based regressor. For the PDR estimation engine, the acc, gyu, mgf and gyu_drift measures are synchronised to calculate the angular velocity of the trajectory. The BLE predictions are leveraged to correct the drifting and misdirection of the PDR estimations. Finally, the BLE and PDR estimations are fused together to generate the final positioning prediction.
Track 6 - Smartphone on vehicle (offsite-online)Team Name SmartLoc
Team members Han Wang
Affiliation TBD
Description TBD Team Name
BJTU-DiDi
Team members Shuli Zhu, Yufei Su, Feng Liu, Haitao Li, Xuan Xiao, Yuqin Jiang, Ruipeng Gao
Affiliation School of Software Engineering, Beijing Jiaotong University
Description Different from the traditional speed estimation method based on attitude calculation and acceleration integration, we designed a supervised deep learning model (Speed DNN Model) for estimating vehicle speed, and estimated the vehicle heading (Bearing Estimation) decoupled from the vehicle speed estimation function in indoor environments. Meanwhile, in outdoor environments, we utilize Extended Kalman Filter (EKF) in order to improve the location of vehicle when GNSS signals and IMU in the smartphone are valid.
The overall workflow of DNN model is divided into two lines: model training and model inference. During model training, we use the provided testing data with high-quality GNSS information to build a training dataset with ground-truth speeds. Among them, we have designed two schemes for the use of inertial data. The first solution is to input the original sampled data (Raw Data) into the deep learning model without additional processing of the inertial data. Team Name
AINS
Team members Zhuang Guangchen
Affiliation Beijing Automation Control Equipment Institute
Description In the first 5 minutes, we use a nonlinear estimation algorithm based on geomagnetic and gravity vectors to obtain the initial attitud. During the validity period of GNSS, we estimate and compensate the installation errors between the IMU, odometer, geomagnetic sensor, and the vehicle. After the degradation of GNSS occurs, the satellite signals are evaluated through a model-based detection method. When the accuracy of satellites cannot meet the requirements of precision, they are promptly excluded from the integrated navigation framework to reduce the negative impact to the navigation system. When the satellite signal fails completely, vehicle navigation and positioning based on fully autonomous sensors are used. A constraint mechanism based on vehicle motion is introduced to further improve the robustness of autonomous navigation and positioning.
Track 7 - 5G CIR (offsite-online)Team Name
WSL Hanayng
Team members Sunwoo Kim, Paulson Eberechukwu N, Minsoo Jeong, Hongseok Jung, Suah Park
Affiliation Hanyang University
Description The WSL team proposes a convolution neural network (CNN)-based stacked autoencoder approach for indoor localization that leverages Channel impulse responses (CIR) and Time-of-arrival (TOA) fingerprints. The proposed approach utilizes a stacked autoencoder network to mitigate the effects of noise and signal fluctuations arising from Non-Line-of-Sight (NLOS) scenarios. The network is trained on pairs of noisy and fluctuating signals and their corresponding clean versions. By minimizing the reconstruction error, the network learns to denoise and restore signals affected by NLOS. The trained stacked autoencoder network effectively reduces noise and mitigates signal fluctuations, enhancing the accuracy and reliability of indoor localization systems in challenging NLOS environments. Team Name
Byr_Trackers
Team members Zhou Heyang, He Jiawei, Song Xudong, Li Shude, Zhang Haoyu, Dong Congrong, Zhang Zhichao, Ding Zhenke, Liu Bingxun, Ma Mingyang, Wang Jizhou
Affiliation Beijing University of Posts and Telecommunications
Description The proposed system consists of two main components: a fingerprint-based positioning system using 5G CSI for non-line-of-sight (NLOS) scenarios and a TOA based positioning system for line-of-sight (LOS) scenarios. In the fingerprint positioning system, we want use a dual-branch CNN network that consists of two parallel sub-networks for amplitude and phase information, respectively. Each sub-network consists of multiple convolutional layers and pooling layers, followed by fully connected layers. The main algorithm of TOA based positioning system is the Nonlinear Least Squares positioning algorithm based on bilateral grid filtering. Team Name
EURECOM
Team members Mohsen AHADI
Affiliation EURECOM
Description Collect ToA from 8 anchors. K-means cluster ToA for LOS/NLOS. Min ToA anchor is reference. Calc LOS TDoA. Extract CIR for TDoA anchors. Sequential NN: input CIR, predict TDoA. Particle Swarm Optimization for Position estimation at the end.
ISCAS
Team members Chang Su, Fusang Zhang, Beihong jin
Affiliation Institute of Software, Chinese Academy of Sciences
Description First, we use machine learning methods that take some hand-crafted features, e.g., the distance corresponding to the highest peak in the CI, as the input to estimate the distance between the transmitter and receiver. Next, we collect the range estimation from anchors with the same burst-id and use a trilateration algorithm to calculate the target's position. Finally, we use filters to smooth the trajectory and resample it to get the final results.
Team Name IOT2US
Team members Yonglei Fan, Qiqi Shu, Zhao Huang, Guangyuan Zhang, Meng Xu, Xijie Xu, Guangxia Yu, Zhao liang Luan, Stefan Poslad
Affiliation Queen Mary, University of London, Peking University
Description We mainly use the acce, gyro and Mega to estimate the step length, step and direction.
it is real time indoor positioning method, we will separate all data into small pieces and use these pieces of data to do the positioning.
For the first positioning, we may use WiFi, or Bluetooth data.
Kalman filter will be used to minimize the error and the cumulative error. But there must be some cumulative error which is hard to deal with. We should think a better way to fix it.
Pedestrian dead reckoning (PDR) has become a research hotspot since it does not require a positioning infrastructure. An integral equation is used in PDR positioning; thus, errors accumulate during long-term operation. To eliminate the accumulated errors in PDR localisation, we proposes a PDR localisation system applied to complex scenarios with multiple buildings and large areas. The system is based on the pedestrian movement behavior recognition algorithm proposed.
Track 8 - 5G ToF (offsite-online)Team Name HHULGD
Team members Huang Ao, Wang Qian, Cui Zhichao, Chen Liang, Cui Yang
Affiliation Hohai University、Army Engineering University of PLA
Description AI algorithm estimation localization We use the AI localization method to establish a mapping scheme between position and measurement through machine learning algorithms and neural networks, thus achieving localization. One of the algorithms is the ELM algorithm, which is currently a type of deep learning. ELM randomly selects the weight of the input layer and the bias of the hidden layer, and the weight of the output layer is calculated analytically based on the (Moore Penrose, MP) generalized Inverse matrix theory by minimizing the Loss function composed of the training error term and the regular term of the weight norm of the output layer. Extreme learning machine has the advantages of less training parameters, fast learning speed and strong generalization ability. In addition, for Track-8, we will also use new deep learning algorithms to help improve model estimation and positioning results. Team Name
GoD
Team members Kai Luo, Ziyao Ma, Yanbiao Gao, Jizhou Wang, Deming Zhu, Yuqi Huo, Tianbao Pan, Xudong Song
Affiliation Beijing University of Posts and Telecommunications
Description According to the provided information of track 8, we proposed a scheme to obtain the high-accuracy position based on the combination of UL-TDOA and AI localization methods.
Our scheme can be divided into four major parts. The first three parts show the preparatory works including preprocessing of training data, machine learning based location estimation and outlier calibration. The last part shows the process of the final test, which includes: 1. Use the trained model to get a tentative result; 2. Outlier detection and calibration; 3. Give the final result. Team Name
ISCAS
Team members Chang Su, Fusang Zhang, Beihong jin
Affiliation Institute of Software, Chinese Academy of Sciences
Description 5G positioning will be part of our lives and it will enable diversified applications in all walks of life. A large number of application scenarios such as the Internet of Vehicles, autonomous driving, smart manufacturing, smart logistics, drones, and asset tracking have higher requirements for positioning capabilities. Excessive errors may lead to poor user experience or other problems. Therefore, we should enhance network positioning technology to improve 5G positioning accuracy. The challenges we face in Track 8 are the features of raw data provided are inaccurate. These features are the results of processing and it’s so hard to infer more information from these features. What’s more, there exists a mixture of LOS paths, weak LOS paths, and NLOS paths. And there are existing timing errors among the receivers in TRPs, called time alignment errors (TAEs). The TAEs of the TRPs are unknown and different in different datasets.
|
