Research On Pedestrian Crossing Intention Recognition Technology Based On Deep Learning

Pedestrians are an important part of the road traffic system and are often in a vulnerable position.Pedestrians crossing the street without zebra crossings,pedestrian crossings and other traffic scenes,as well as grabbing red lights to cross the street,due to driving blind spots and other factors lead to frequent traffic accidents,research on pedestrian crossing intention identification can effectively predict pedestrian crossing behavior,avoid traffic accidents in advance,and effectively ensure the safety of pedestrians crossing the street.Aiming at the problems of high false detection and leakage rate of small target pedestrians,low tracking accuracy of motion trajectory,and difficulty in identifying pedestrian movement intentions in complex traffic scenarios,the relevant information such as pedestrian movement trajectory and human posture is fully excavated to effectively express pedestrian crossing intentions.Based on the theories and methods of deep learning,the methods of pedestrian target detection,trajectory tracking,attitude discrimination and crossing intention recognition in complex traffic scenarios are mainly studied,and the main contents are as follows:(1)A pedestrian detection method based on attention mechanism and multi-scale feature fusion is proposed.Aiming at the problems of high pedestrian false detection rate,low dense pedestrian detection accuracy,and high pedestrian missed detection rate of long-range small targets,a pedestrian detection method based on attention and multi-scale Feature Fusion-You Only Look Once(AMFF-YOLO)is proposed.Firstly,the attention mechanism module is introduced in the feature extraction network,which enhances the key feature information and suppresses the background information.Secondly,in the feature fusion network,the Multi-scale Feature Fusion Method(MFFM)is embedded,which enables the effective fusion of feature information between different scales.Then,a large-scale detection layer is added to the detection network to enhance the feature detection performance of pedestrians with small targets in the long range.Finally,comparative experiments,generalization experiments and other verification experiments were carried out on the public BDD100 K and Crowd Human pedestrian datasets.Experiments show that the detection accuracy of the proposed method reaches 71.0% and 86.2%,and the recall rate reaches 62.6% and 54.3%,respectively.(2)A pedestrian multi-target tracking method based on full scale is proposed.Aiming at the problems of low multi-target pedestrian tracking accuracy and poor matching of pedestrian appearance features,a full-scale pedestrian multi-target tracking method is proposed.Firstly,the AMFF-YOLO pedestrian detection method is embedded in the tracking algorithm,and the accuracy of the pedestrian multi-target tracking algorithm is affected by the detector accuracy,which improves the detector detection accuracy and then improves the tracking accuracy of the tracking algorithm.Secondly,the full-scale pedestrian appearance feature extraction model is used to distinguish the similarity between the same pedestrian in different frames and the difference between the appearance features of different pedestrians in the same frame,so as to reduce the situation of pedestrian target matching errors.Then,the GIOU matching module is added,and the minimum bounding box between the two frames is used to determine which prediction frame and the detection frame match the detection frame best when there is no overlap in the position of the trajectory prediction box.Finally,the model is experimentally verified in the public MOT17 pedestrian dataset,and the final MOTA value is 66.3%,and the IDSW number is 593,and the experimental results show that the full-scale pedestrian multi-target tracking method can better track multiple pedestrians continuously.(3)A pedestrian crossing intention recognition method based on LSTM is proposed.Aiming at the problems of low recognition accuracy in pedestrian crossing intention recognition network,pedestrian crossing intention recognition is regarded as a time series modeling prediction problem,and a pedestrian crossing intention recognition method based on LSTM is proposed.Firstly,the Open pose pose recognition network is used to extract the key point information of the human body,and the key point information is optimally matched to fit the attitude characteristics of the traveler.Secondly,the full-scale pedestrian multi-target tracking network is used to track pedestrians and extract their trajectory information.Then,the attitude characteristics and motion trajectory information of pedestrians are input into the LSTM model,and the information is fully utilized to achieve the purpose of accurate identification.Finally,an experimental verification on the JAAD dataset achieves an 85.2% recognition accuracy,which shows that the model can accurately identify the pedestrian’s crossing intention.