Pedestrian Detection And Re-identification Based On Deep Learning Technology Research And Application

Person re-identification refers to the technology of using computer vision technology for pedestrian matching based on the video collected by multi-camera network with non-overlapping field of view,that is,the algorithm automatically confirms whether the pedestrian target captured by cameras at different positions and different times is the same person.Person re-identification is of great application value in public security and criminal investigation and image retrieval,so it is of great research significance.Pedestrain detection,which is a technology to localize pedestrian in images,is a fundamental of person re-identification.In this paper,pedestrian detection algorithm based on RetinaNet and person re-identification based on Meta-SR combined with ID-discriminative embedding network are studied.(1)The situation that speed and accuracy cannot be obtained at the same time has always been the bottleneck of the target detection task.Single-stage detection methods such as Yolo are fast in detection speed,but have problems such as low detection accuracy,low recall rate and poor detection effect on small targets.Faster r-cnn and other two-stage methods can solve the problems of low detection accuracy and poor detection effect of single-stage detection method,but the detection speed is slow.RetinaNet studied that the main reason why the single-stage detector is less accurate than the two-stage detector is that the single-stage detector has serious "category imbalance" problem,which leads to the failure of classifier training.In addition,a focus loss function was proposed to solve the "category imbalance" caused by excessive background.Based on RetinaNet,this paper improves the application characteristics of pedestrian detection in the following three aspects:1)Improves the network structure and changes the feature extraction method;2)Channel attention module is introduced better to learn the features that are effective for pedestrian detection;3)Design pre-selection box according to pedestrian characteristics to improve the effect of pedestrian detection.The experiment proves that the improved RetinaNet in this paper has obvious performance improvement for pedestrian detection.(2)Most of the existing person re-identification method again assuming pedestrian images with uniform size,and have a high enough resolution,they are usually all image is normalized to the same size,then input to the re-identification system,while ignoring the real scenario query image is usually a high resolution,and the image resolution in the candidate pedestrian database is generally low.As a matter of fact,due to the blurred image and low resolution of the monitoring video,Various resolutions coexist and scale mismatches of pedestrian images have been existing in the real world.In order to solve this problem,this paper studies the person re-identification technology,based on Meta-SR by fusion of person re-identification network and super resolution network to identify the input to the resolution of the images before re-identification module for pedestrians to properly,the problem of coexistence of different resolutions and scale mismatch in person re-identification system is solved effectively.