Research On Multiple Pedestrian Tracking Based On Deep Learning

In most other academic studies of multi-target pedestrian tracking in the past,the positions and trajectories of pedestrians are only in two-dimensional image coordinates,which are still very different from the positions and trajectories of pedestrians in real three-dimensional space.Therefore,this paper propose a new multi-camera-based joint calibration algorithm based on the traditional camera calibration method.We establish a unified threedimensional world coordinate system for the multiple cameras in the tracking system,so that the position of pedestrians in three-dimensional space can be effectively and accurately tracked.In addition,the existing algorithms are too time-consuming and difficult to implement in real time.Inspired by this,this paper proposes a multi-target pedestrian tracking system running in real time on multiple cameras.This paper have three contributions.Firstly,we proposed a new multicamera-based joint calibration algorithm based on the traditional camera calibration method.We use principal component analysis to reduce Z axis error and use multi-camera parameter transfer to reduce origin calibration errors.We establish a unified three-dimensional world coordinate system for the multiple cameras in the tracking system,so that the position of pedestrians in three-dimensional space can be effectively and accurately tracked.And based on that,this paper proposes and implements a lightweight 2D and3 D joint tracking algorithm.This paper makes full use of the position and motion information of the 3D trajectory to make the 3D association more accurate.We removed the process of finding the optimal path to make the algorithm more lightweight.And we back-project 3D trajectories to improve the accuracy of 2D tracking.Secondly,in order to achieve real-time tracking this paper proposes a global-to-local tracking-by-detection algorithm.We use global detection at key frames while we use motion prediction and local detection at non-key frames.Because the local detector makes full use of the prior information of the target position in time series,it can use a simple network to track the target well.The algorithm enables multi-target tracking improving time performance without much decrease of accuracy.Thirdly,this paper integrates all algorithms mentioned above to implement a robust multi-camera-based real-time multi-target pedestrian tracking system.And we build a pedestrian head and shoulder detection dataset with more than12,000 images to implement this system.Comprehensive experiments well demonstrate the outstanding performance of the proposed method in terms of efficiency and accuracy in multitarget 3d trajectory tracking tasks.The experiments designed in this paper include ablation study in accuracy and real-time analysis and analysing the accuracy and real-time performance of the integrated system.The first experiment is comparison of the accuracy of traditional calibration methods and our joint calibration methods in a laboratory environment.The second one is to demonstrate the performance of motion prediction models on GOT-10 k datasets.The third one is to evaluate global and local detectors on self-built pedestrian detection datasets.The last one is to evaluate overall performance of the whole system on PETS-09 tracking dataset.