Research On Pedestrian Detection And Re-identification In Video Surveillance

With the development of computer technology and sensor technology applied in the field of intelligence prevent,a large amount of video monitoring data has been produced,the effective analysis of massive video data has important theoretical value and practical significance for military and civilian applications such as ordnance automation and unmanned driving.Pedestrian Re-Identification(Re ID)is a hot topic in many research fields,which can realize intelligent retrieval of target pedestrians under complex environment.Traditional Re ID technology can effectively realize pedestrian re-recognition based on handdesigned visual features,and has made important progress.However,most of the methods are difficult to obtain a good similarity measure,which restricts the improvement of Re ID model performance.With the development of artificial intelligence technology,data-driven deep learning methods can better express pedestrian features and provide new methods and ideas for solving Re ID problems.However,Re ID based on artificial intelligence technology is still difficult to meet the needs of realistic complex scenes,especially the retrieval efficiency and quality of specific suspect targets need to be improved.Based on deep learning,this paper conducts research on key issues such as pedestrian occlusion,pedestrian posture change and human body deformation in video surveillance.The main research contents and results are as follows:1.In order to address the overfitting problem of deep learning framework for Re ID,an DYOLOv3 pedestrian detection model was proposed.By reducing the number of network structure layers,it can detect pedestrians on four feature maps of different scales,which effectively improved the pedestrian feature modeling ability of the model.Experimental results show that the model can effectively improve the detection accuracy of pedestrian targets.2.In order to address the problem of misdetection and omission in surveillance video scene caused by pedestrian partial occlusion,posture change and human body deformation,a Global Joint Local Batch Normalized Res Net(GLBNRes Net)is proposed to extract the image features of pedestrians to be retrieved in surveillance video scenes.By integrating global and local features and constructing the similarity measurement function,this method can reduce the deviation between the prediction frame and the real frame of the same target,and reduce the error caused by human deformation and occlusion,so as to effectively reduce the false detection rate and missed detection rate.Experimental results show that the proposed GLBNRes Net can realize pedestrian detection and Re ID in complex scenes,and has a good performance.3.In order to address the problem of modeling the depth feature of Re ID,a measurement model combining Quadruplet loss,Central loss and Identification(ID)loss was proposed,and a DYOLOv3-GLBNRes Net Re ID system was constructed.First,the DYOLOv3 network is used to detect pedestrians in the video.Then,the GLBNRes Net network is used for feature extraction.Finally,the re-identification results of the given target pedestrians are output.The experimental results show that the DYOLOv3-GLBNRes Net Re ID system can realize pedestrian re-identification in complex scenes,and has high detection accuracy and strong robustness.