Prediction Of Pedestrian Motion Trajectory Based On Deep Learning

With the increasing use of motor vehicles,the incidence of traffic accidents is also increasing.The inability of vehicle drivers to respond in emergency situations is one of the main factors leading to traffic accidents.Assisted driving technology can sense the surrounding environment during the driving process of the vehicle and help the driver make judgments,thereby improving driving safety.Pedestrian trajectory prediction is one of the important components of environmental sensing,and the trajectory prediction method based on deep learning is more generalized and can better meet the actual application needs,but there are still the following problems:1)Trajectory acquisition time is inefficient,Poor performance under occlusion conditions;2)Trajectory prediction has high time complexity and harsh application conditions.In view of the above problems,this thesis aims to improve the speed and accuracy of road pedestrian trajectory prediction,and designs a road pedestrian trajectory acquisition and prediction algorithm based on deep learning.The main contents of this thesis are as follows:1)Aiming at the problems of low efficiency of trajectory acquisition time and poor effect under occlusion conditions,based on DeepSORT,YOLOv5m is first used to replace Faster-RCNN for detection,which improves detection efficiency and realizes real-time tracking.Secondly,the loss term L_{Re p Box}in Repulsion Loss is added in YOLOv5m,and GIo U Loss is replaced by CIo U Loss,and the DIo U-NMS strategy is applied to improve the tracking accuracy under occlusion.On the City Persons dataset,YOLOv5m is tested for time efficiency and occlusion detection,and YOLO5m is combined with DeepSORT to perform tracking comparison experiments on the MOT16 dataset.The results show that the trajectory acquisition algorithm designed in this thesis shows better tracking rate and accuracy,and can better deal with occluded targets.And on the basis of LSTM,a gated unit is added to obtain an M-LSTM structure network,and M-LSTM is used to replace the convolutional neural network in PMTP for data feature learning,which is used for pedestrian motion speed prediction.2)Aiming at the problems of high time complexity and harsh application conditions of trajectory prediction,a local weighted regression algorithm is designed according to the principle of PMTP algorithm,which is used for the prediction of pedestrian movement trends;On the datasets collected by the FPL algorithm,Caltech Pedestrian datasets and NGSIM datasets,the prediction accuracy of the algorithm designed in this thesis is compared.Experiments show that the trajectory prediction algorithm designed in this thesis has higher accuracy than the FPL and PMTP algorithms for short-term prediction,and the FPS index can better meet the real-time prediction requirements.3)Based on the designed algorithm,an assisted driving early warning prototype system is developed.The system includes a pedestrian trajectory tracking and prediction module.The pedestrian trajectory tracking module is used to track the trajectory of multiple pedestrians in the image captured by the camera;The pedestrian trajectory prediction module will first determine the early warning area,and if the pedestrian trajectory predicted by the algorithm comes into contact with the early warning area,an alarm will be triggered.At the same time,this thesis also collects the driving data set of the driver’s perspective,and tests the system to verify the effectiveness of the designed pedestrian trajectory tracking and prediction algorithm.