Research On Pedestrian Tracking And Trajectory Prediction On Urban Road Based On Deep Learning

In recent years,the breakthrough of data-driven deep learning technology has promoted the application of autonomous driving technology in limited urban areas.However,the pedestrian priority strategy adopted by self-driving vehicles for safety reasons will cause them to stagnate in the high-density urban road environment,causing traffic congestion.Accurately perceiving pedestrians and predicting the pedestrian's future trajectory is the key to solving this problem.This paper focuses on the problems of pedestrian detection,pedestrian tracking and pedestrian trajectory prediction in the autonomous driving environment perception scheme.Based on deep learning methods and visual sensors,object recognition,location acquisition and trajectory prediction research are carried out for pedestrians on urban unsignal controlled sections,The main research contents are as follows:(1)A one-stage target detection model based on convolutional neural network is built.Current autonomous driving technology requires the target detection model to have good light change robustness and scene adaptability.In response to this demand,a pedestrian detection model is constructed based on Yolov3,a one-stage target detection algoritlun.USing K-means clustering the VOC2007 data set,9 more universal pedestrian detection frame sizes are obtained,and on the basis of the MS COCO data set,the INRIA Person data set and ExDark data set are added to expand the pedestrian samples of the day and night scenes.The final trained model is robust to lighting changes on the CUHK square,CPCS,KITTI,ExDark and NightOwls test sets,with good scene generalization performance and high accuracy and real-time performance.(2)A pedestrian tracking model that introduces a global attention mechanism is proposed.The current multi-target tracking model of TBD(Track by Detection)mode,due to the poor internal cascading matching effect,causes frequent target ID switching during the tracking process.To solve this problem,this paper proposes to introduce the global attention mechanism into the apparent feature extraction process of cascade matching.The proposed model is based on SORT,and the Rga(Global Attention)module is added to the apparent feature extraction network ResNet50 to improve the characterization ability of the network output.The proposed model is robust to illumination changes on the autonomous driving data set KITTI of daytime scenes and NightOwls of nighttime scenes,and the generalization of scenes and accuracy are good,and on the MOT 16 benchmark,compared with the four methods of EAMIT,POI,SORT and DeepSORT,the method in this paper has improved some accuracy indicators such as ID switching and can realize real-time tracking on a single graphics card.(3)A LSTM pedestrian trajectory prediction model with state information sharing is proposed.The original LSTM unit structure only predicts the trajectory based on the state of the target to be predicted,so the prediction result cannot reflect the group and mutual avoidance of pedestrian trajectories.To solve this problem,a LSTM unit structure with state information sharing is proposed.The proposed structure considers the state information of other objects within the specified distance of the target to be predicted,and j oins the construction process of the LSTM internal long-term memory unit in the form of a weighted sum of state information.Based on the proposed structure,an LSTM trajectory prediction model with state information sharing is designed.This model can achieve real-time trajectory prediction on a single graphics card,and on the ETH and UCY benchmarks,compared with SF,IGP,LSTM and the Social-LSTM method that also considers the interaction of pedestrians with each other,the lowest average displacement error is achieved,and the final displacement error is kept low.(4)Use self-collected campus road scene data set to verify the performance of the three models in this paper.The target detection model built in this paper and the pedestrian tracking model proposed in this paper have been further verified on self-collected data sets for their ability to cope with changes in illumination and scenes.The three models maintain high accuracy and real-time performance in actual automatic driving scenarios,and can meet the needs of pedestrian detection,pedestrian tracking,and pedestrian trajectory prediction tasks in the field of automatic driving.