Research On Pedestrian Detection Algorithm Of Bridge Crane System Based On Deep Learning

Pedestrian detection and tracking technology in video sequences is an important field in computer vision.For video surveillance in the bridge crane scene,there are some special problems in pedestrian detection and tracking.Specifically,the working scene of a general bridge crane is an outdoor environment,which is easily affected by light,weather,and so on.On the other hand,motion detection is a time-consuming operation in the monitoring system.Therefore,the selected detection algorithm needs to be able to run quickly and robustly to ensure the real-time performance of the system.This paper proposes an adaptive pedestrian detection method based on visual attention mechanism.First of all,in view of the very important interaction between feature-selective and spatial attention in the human attention mechanism,this paper proposes to realize the adaptive adjustment of the receptive field size of neurons through the additive effects of feature-selective and spatial attention.Thus,an adaptive feature extraction network based on visual attention mechanism is designed.In addition,because the pixel-by-pixel feature fusion mechanism can fuse the feature information of corresponding pixels in the corresponding channels of the feature map,the operation does not change the number of channels of the feature map,but can increase the amount of information in the output feature map in each dimension.Obviously,the classification of the network is beneficial.Following this principle,this paper introduces a pixel-by-pixel feature fusion mechanism in the multi-scale prediction of the original YOLOv3 detection algorithm.In this paper,a multi-scale pedestrian tracking method CSSA-Siam based on the two-branch Siamese fully convolutional network is proposed.Research on pedestrian tracking related issues in bridge crane systems found that pedestrians have the characteristics of scale changes.Based on this observation,a new feature extraction network based on multi-scale fusion is proposed.Then,in order to enhance the ability of the semantic branch in the SA-Siam network to learn semantic features,the CS attention module is used to replace the channel attention module in the semantic branch of the SA-Siam network.Finally,the proposed pedestrian detection method is analyzed on PASCAL VOC 2007,PASCAL VOC 2012,USC-A,USC-C,INRIA datasets and crane background image datas,and experimental results show that the proposed pedestrian detection method has a higher detection accuracy while maintaining better detection real-time performance.Experiments on OTB100 datasets verify the effectiveness of the proposed pedestrian tracking method,and its tracking accuracy and tracking speed reach 85.9% and 34 FPS,respectively.The detection and tracking algorithm designed in this paper can effectively monitor pedestrians in the video,so as to provide safety guarantee for the safe operation of the crane and achieve the goal of safe and efficient production.