Pedestrian Automatic Detection And Re-Identification In Complex Scenes

In recent years,with the increasingly applications of intelligent video surveillance,intelligent transportation,vehicle-assisted driving systems,and intelligent robots,more and more tasks need to be done with the help of computer vision technologies.In particular,pedestrian detection and person re-identification have become the hot research topics in computer vision.They have been widely studied by researchers in related fields and developed rapidly.However,due to the complexity of the surveillance,including the variability of pedestrian posture and the ubiquity of pedestrian occlusion in practical application scenarios,there are still many challenges and difficulties in pedestrian detection and person re-identification in complex scenes.First,pedestrians have problems with low resolution,large scale differences,and severe occlusion in complex scenes.Based on the Faster R-CNN pedestrian detection algorithm,this thesis uses Precise RoI Pooling and multi-scale receptive field network to improve the algorithm.On the one hand,they increase the diversity of the receptive field scale of the network features,and on the other hand,improve the accuracy of the feature extracted by Region of Interest Pooling.The improved algorithm has better adaptability to pedestrian detection in complex scenes.Secondly,this thesis proposes a two-channel network model based on Faster R-CNN.For pedestrian detection in low light environments,thermal images have more information than RGB images.The fusion of RGB image and thermal image feature is realized by a two-channel network,which can complement the information between the two types of features,and increases the diversity of pedestrian features.The improved algorithm is more robust to ambient light and improves the performance of the pedestrian detection algorithm in low light conditions.Thirdly,considering the changes in the appearance of pedestrians and the obvious changes in the perspective of pedestrian images,based on the PCB person re-identification model,this thesis proposes a PCB-DB model with joint feature diversity and weighting local feature.It combines the global and local features of pedestrian images and enriches the effective information of pedestrian features.This model can distinguish multiple local features by weighting local feature and guide the PCB-DB model of training by jointing Cross Entropy Loss and Batch Hard Triplet Loss.The improved model can better adapt to the person re-identification task in complex scenes.Finally,an automatic pedestrian detection and re-identification software system for video surveillance scenarios is built.The system can process monitoring video data in multiple different scenarios and implement pedestrian detection and re-identification across scenes.