Research On Pedestrian Tracking And Trajectory Prediction Under Abnormal Environment Based On Deep Learning

Pedestrian detection,tracking and trajectory prediction in abnormal environment are important issues in pedestrian traffic safety research.As pedestrians are more prone to safety accidents in abnormal environment,it is of great significance to study pedestrians in abnormal environment to ensure pedestrian traffic safety and reduce risk losses compared with normal pedestrian research.In view of the few systematic studies on pedestrian detection,tracking and trajectory prediction under abnormal environment and the shortcomings in research methods,this paper focuses on pedestrian detection,tracking and trajectory prediction in abnormal environment based on deep learning.The work done in this paper mainly includes the following six parts:1.Focusing on the development trend of pedestrian detection,tracking and trajectory prediction in abnormal environment,the paper elaborates the research status of pedestrian detection,pedestrian tracking,pedestrian trajectory prediction and pedestrian traffic flow theory,and analyses the necessity of pedestrian research in abnormal environment based on deep learning.2.Two kinds of advanced target detection Frameworks-YOLOV3,SSD are compared in terms of design concept,feature extraction network and loss function.Pedestrian detection results in abnormal environment are compared by retraining the framework with self-made datasets-APD and public datasets-VOC.The results show that the detection accuracy of YOLOV3-VOC+APD for pedestrians is 0.88,which is better than the baseline model.Through visual analysis of inspection results through examples,it is found that YOLOV3-VOC+APD can effectively identify ambiguous targets,small targets and edge targets in abnormal environment.3.Based on YOLOV3-VOC+APD bottom detection framework,a pedestrian tracking algorithm with deep feature is proposed,which is compared with the baseline Sort algorithm on public datasets,and proves that the tracking frame generated by the tracking algorithm with deep feature can match the detection frame well and solve the problem of ID frequent switching caused by long-term shielding between pedestrians.4.In order to compare the accuracy of pedestrian trajectory extraction by the tracking algorithm,considering that the trajectory extracted by the tracking algorithm is an incomplete sequence trajectory,a dynamic time warping(DTW)algorithm is used to align the trajectory extracted by the tracking algorithm with the trajectory data of manual markers in time dimension and to evaluate the trajectory similarity,taking three pedestrian trajectories in abnormal environment as an example to test the experimental performance.The results show that the trajectory extracted by the algorithm is highly similar to the trajectory data of manual marking,in which the similarity distance of trajectory in both longitudinal and transverse directions is less than 0.03.5.An automatic collection method of pedestrian traffic flow parameters in an abnormal environment is proposed,which extends the trajectory tracking algorithm incorporating deep characteristics.The method mainly includes:flow collection based on cross-line rules,density collection based on area of interest and density monitoring based on graph superposition,and speed calculation based on interpolation of pedestrian trajectory.The experimental results show that the maximum pedestrian flow rate of concerned section under abnormal environment is 1ped/(m.s),corresponding to service level D;the maximum instantaneous density is 0.29ped/m~2,and pedestrians have relatively free space in this abnormal environment;the average speed is 2.6m/s,and the speed between pedestrians is different,with standard deviation of 1.1m/s.6.A pedestrian trajectory prediction model Seq2Seq-GAN considering pedestrian interaction is proposed for abnormal pedestrian trajectory data.Seq2Seq-GAN is compared with several trajectory prediction baseline models in generalization performance on open datasets.The experimental results show that the average displacement error of Seq2Seq-GAN is 0.59m and the final displacement error is 1.19m,which proves the validity,applicability and advancement of Seq2Seq-GAN for pedestrian trajectory prediction.Visualization of pedestrian trajectory prediction results in abnormal environment is carried out by combining the model.The results show that the average displacement error of the trajectory prediction results is less than 1m.Although there is some deviation between the predicted trajectory and the real trajectory,the trend is basically consistent and the predicted trajectory results conform to social norms.In addition,considering the impact of static targets can effectively improve the forecast accuracy of the model.