Research On Pedestrian Detection And Behavior Analysis Algorithm Of Crosswalk Based On Deep Learning

In the mixed traffic of cities,due to the rapid development of urban motorization,cars occupy the strong position,and the position of people is weaker.Since people cross the road generally in the crosswalk area,the crosswalk area often becomes the most frequent area where people interact with the vehicle.The location and crossing mode of people in the crosswalk area has become a key factor in the judgment of regional traffic safety.Therefore,this thesis starts from the two aspects of pedestrian detection and behavior analysis based on video surveillance,and studies the corresponding intelligent algorithms based on deep learning.The main work of the thesis includes:(1)A backbone network for object detection is proposed and applied to the onestage target detection network SSD for the pedestrian detection in crosswalk area.Firstly,an in-depth analysis of the one-stage object detection algorithm SSD based on deep learning is carried out.Then,the backbone network used by the detection network was discussed,and the existing backbone network DarkNet-53 was initially improved,and the initial improvement network DarkNet-60 was obtained.Then,a network structure Dense RFB(receptive field block)for enhancing the network receptive field is proposed,and the structure is added to the initially improved backbone network to obtain the DRFNet(dense receptive field network).Finally,the backbone Network DRFNet is used in SSDs to improve the detection performance of SSDs on zebra crossings.(2)A fast pedestrian detection algorithm based on the detection algorithm CenterNet is proposed.Firstly,CenterNet,which is anchor-free algorithm,is chosen to replace the anchor-based detection algorithm SSD to avoid the cumbersome tuning problem in the actual application.Then,the CenterNet algorithm is improved in two aspects including network structure and loss function.In terms of network structure,the FRFEM structure was proposed and RFRNet was constructed.RFRNet can effectively enhance the distribution of the receptive field of the backbone network and improve network performance through additional connections.In terms of the loss function,the Cener-IoU Loss with stronger numerical stability and easier optimization is proposed.Finally,by taking the advantage of its anchor-free characteristics,the transfer learning on the large-scale pedestrian detection dataset is applied to it,enabling the algorithm to quickly and accurately detect crosswalk pedestrians.(3)A behavior analysis algorithm based on timing information of human skeleton was proposed to classify the pedestrian traffic pattern in the crosswalk area.First,the time series images of human body in video monitoring were obtained by using multiple object tracking algorithm.Then the pose estimation algorithm is used to extract the pedestrian skeleton information from the time series image.Then,the RGB image is formed by mapping the skeleton information of pedestrians,and the behavior recognition model is constructed by using the classification network to classify the mapped action image,and finally the crossing mode of pedestrians in the crosswalk area is judged.