Research On Multi-scale Pedestrian Detection In Complex Backgroun

Pedestrian detection is an indispensable part in the field of computer.As the most basic section in tasks,such as automatic driving,monitoring of pedestrian flow,object re-identification,pedestrian tracking,behavior analysis,etc.,pedestrian detection has been widely used in intelligent transportation systems and intelligent monitoring systems.In recent years,because of high detection accuracy and fast detection speed,the pedestrian detection algorithm based on deep learning has gradually replaced the traditional pedestrian detection algorithm.However,because the actual application scenes are complex and random,there are still many problems in the pedestrian detection algorithm based on deep learning when detect the different scales pedestrian in complex background.In this paper,the problems in the practical application are studied,for example,pedestrian occlusion,insufficient light,and missing small-scale pedestrian,etc.The main research contents can be summarized as follows:1.In order to reduce the impact of environmental factors on pedestrian detection algorithm in practical application,pedestrian dataset including 3270 pictures are made,taking scale,angle,background and light as the collection elements.Using the self-made dataset,the classic pedestrian detection algorithm,YOLOv4 and YOLOv5(three models),is tested.According to the experimental results,YOLOv5 s pedestrian detection algorithm with better robustness and detection effect is selected as the basic algorithm,and the problems of YOLOv5 s about detecting pedestrians in complex background are analyzed.2.An improved YOLOv5 s algorithm is proposed to solving false detection and missing detection for multi-scale pedestrians under complex situations,such as occlusion and insufficient light.Firstly,to make the model has higher detection accuracy when detecting pedestrian with complex background,the robustness of model is enhanced by designing a multi-method data enhancement for training process.Secondly,the CBAM attention mechanism is introduced into the shallow network,for having the weights of channel and location dimensions,while ignoring the feature map with unimportant information.Finally,the improved YOLOv5 s algorithm is trained on the self-made dataset,and 81.08% of the AP value is obtained,which is 3.8% higher than the original algorithm in the complex background,which verifies the effectiveness of the algorithm.3.In view of the slow detection speed of YOLOv5 s algorithm in the embedded system development,a lightened YOLOV5 s algorithm is proposed to meet the real-time detection.Firstly,the main reason for the slow speed is analyzed through the visualization of feature map,that is,there are some redundant features in the feature extraction process.The redundant features can waste a lot of computing resources.Secondly,the lightened YOLOv5 s network model is lightened by introducing the Ghost module and reducing the number of channels.Compared with YOLOv5 s algorithm,the AP value decreases by 0.88%,but the parameter quantity of the model is reduced to 1/5 of YOLOv5 s,greatly accelerating the prediction speed of the algorithm.Finally,the lightweight network model is deployed on the Jetson Nano embedded development board.And the prediction speed of the network is further improved through Tensor RT,realizing the real-time detection on the mobile devices,and effectively alleviate the error detection and missed detection in backgrounds.