Research On Detection Method Of Pedestrian Shoes Under Surveillance Video

"Monitoring + shoe print" is an important technique in the criminal investigation process of the public security organs.This technique combines the shoe print traces at the crime scene with the surveillance video around the scene,so as to associate the shoe prints with personal information,and provide clues for the investigation and solving of the case.However,its application is limited due to the low automation degree of this technique and the timeconsuming and labor-intensive screening of videos.Hence,a corresponding shoe detection algorithm is urgently needed to improve the efficiency of video screening and save time in solving the case.At present,there are two main problems: one is the lack of a publicly available database for the scenario;the other is the difficulty to detect the pedestrian shoes in the surveillance video considering it is a small target.With regards to above problems,this paper conducts the research in the following four parts.First,a database of shoe patterns under the surveillance video of walking pedestrians is built.Firstly,a scene is set up in the laboratory to simulate various road conditions.Secondly,343 different types of shoes are selected to meet the requirements of actual cases.Then volunteers are invited to assist the video collection.Finally,a database of 20,461 images is formed through decoding,framing,filtering and labeling the video data,providing data support for subsequent experiments.Second,considering the characteristic of detection problems about walking pedestrians under monitoring,an automatic detection algorithm is proposed based on YOLOv4.Firstly,the K-means clustering algorithm is used to determine the scale and quantity of the anchor box.Secondly,the appropriate detection layer is selected to strengthen the learning of shoe features.Then more feature information is obtained through multi-scale feature fusion.Finally,the adjusted spatial pyramid pooling module is transferred into the improved network to enhance the learning performance of the model.The size of the YOLOv4＿shoe algorithm proposed in this paper is 39.56 MB,with an average accuracy of 99.43%.Third,considering the real-time detection requirements of actual cases,a lightweight detection algorithm based on YOLOv4＿Tiny is proposed.Firstly,depthwise separable convolution is used to reduce network parameters and improve detection efficiency,then the original model is compressed and the depth of the detection layer is reduced to enhance the performance of small target detection.Secondly,the Focus structure is used to achieve downsampling with less distortion and DS＿SPP＿Block and DS＿SSH structures are constructed to obtain more context information to enhance the network feature extraction ability.The attention mechanism is also introduced to strengthen the positioning of the anchor box,and the Swish activation function is used to deal with the loss of accuracy caused by the reduction of model complexity.The YOLOv4＿Tiny＿shoe algorithm proposed in this paper is characterized by a size of 1.80 MB,a parameter value of 0.44 M,an average accuracy of 98.11%,and an FPS value of 90.47,which meets the requirements of actual cases.Fourth,the calculation steps of shoes’ anchor box are simplified and a detection algorithm based on Center Net is proposed.Firstly,the downsampling rate of the backbone network is reduced,and the convolution module of the downsampling layer is increased by 4 times and 8times to strengthen the feature extraction of small targets.Secondly,the transposed convolution structure is increased to obtain feature maps with a higher resolution and strengthen the positioning of the bounding box.Then a sharing mechanism of lateral feature information is designed to integrate the shallow features of the backbone network with the high-resolution features obtained by upsampling,so as to strengthen the descriptiveness of the feature information for the detection target.Finally,the MRF＿Block structure is designed to apply different scales of dilated convolutions and maxpooling to obtain multi-scale receptive field information.The Center Net＿shoe algorithm proposed in this paper is 24.03 MB in size,with an average accuracy of 98.10%,and the detection box is more suitable for shoe detection with a more ideal detection result.In this paper,a database with various types and sufficient data is constructed for detection tasks of pedestrians’ shoes in the surveillance videos,and three different detection algorithms are proposed to meet the detection requirements in different scenarios,which will help criminal investigators quickly screen videos and improve the automation degree of the "monitoring +shoe print" technique.