Moving Pedestrian Detection Based On Fusion Of Foreground Detection And Deep Learning

Because of its wide application prospects in the field of intelligent security,moving pedestrian detection technology has become one of the research hotspots in computer vision.At present,moving pedestrian detection has been applied well in simple scenes with fixed shooting angles.However,the actual monitoring scenarios vary,and pedestrians' occlusion,overlap,complex environment,and false alarm targets around them all have a great impact on the results.Therefore,it is difficult to design a moving pedestrian detection algorithm with real-time performance,high accuracy and recall rate.In view of this problem,this paper carries out relevant research on the basis of full analysis of existing methods.The main work of the paper is summarized as follows:(1)This thesis introduces the theory and tools of deep learning,and takes YOLOv3-Tiny network as the prototype,studies a new pedestrian detection algorithm based on improved YOLOv3-Tiny network.Firstly,according to the characteristics of the foreground detection bounding box,the pedestrian as a whole is determined as the detection target.Secondly,the process of the algorithm and pedestrian prediction are introduced.An optimization scheme is proposed aiming at the shortcomings of YOLOv3-Tiny network structure: The max pooling layer in backbone network is replaced by 3×3 convolution layer,which reduces the loss of features while compressing the feature map;A feature pyramid branch is introduced in the shallower layer and combined with the deep feature in order to improve the detection accuracy.The experimental results show the algorithm can effectively improve detection accuracy compared with original YOLOv3-Tiny network based algorithm.(2)This thesis introduces the arithmetic principle,advantages and disadvantages of traditional background modeling method and inter-frame difference method.For the requirement of target extraction speed in subsequent moving pedestrian detection,a moving object detection algorithm based on background difference method is introduced.The results show that the introduced method has certain advantages in detection speed compared with traditional foreground detection algorithm.(3)For the problem of moving pedestrian detection,a fusion detection algorithm is proposed,which combines the bounding boxes of the two detection algorithms mentioned above.Firstly,the characteristics of the detection bounding boxes of the two detection algorithms are analyzed: Foreground detection usually has a relatively low accuracy in marking pedestrian information,but it can effectively remove the static false alarm targets;Deep network detection methods are usually difficult to obtain both high accuracy and recall rates,but it can flexibly adjust the level of both by setting confidence threshold.Secondly,the feasibility of fusing two kinds of bounding boxes is discussed,and a fusion detection algorithm is proposed.The experimental results show the algorithm can improve the accuracy and recall rate,and reduce the false alarm rate compared with the improved YOLOv3-Tiny network based detection algorithm.