Design Of Video Retrieval System For Pedestrian Monitoring In Building Space

With the development of the Internet and the advancement of the Skynet Project,the number of surveillance cameras has grown rapidly.By 2019,770 million surveillance cameras were used in various building spaces worldwide.By 2020,the number of cameras installed in China is expected to reach 626 million,accounting for the largest market share in the world.A large amount of image data were produced by the whole city coverage of the surveillance cameras.How to analyze and process these video data efficiently,and retrieve the required video fragments to meet user needs is an urgent problem.At present,the analysis method of surveillance video is primitive,which relies heavily on manual screening,which is inefficient and costly.Monitoring personnel on long-term duty are prone to attention fatigue,resulting in safety hazards and low accuracy risk.The traditional pedestrian retrieval system generally adopts face recognition technology,which is based on the information of human face feature.However,but the real environment is complex and diverse.The back of the head,side of the face and external environmental factors such as shielding and angle will cause abnormal face recognition from the perspective of actual monitoring.And the pixel of the surveillance camera is not high,it is difficult to capture a clear face from the perspective camera.Therefore,considering the existing problems and difficulties of building space surveillance video retrieval process,based on traditional content-based image retrieval,the paper proposes the design and implementation of pedestrian surveillance video retrieval system based on HSV color space and ResNet50 network two-stage retrieval algorithm,using characteristics of deep learning technology such as strong learning ability,portability and adaptability.In the paper,first of all,from the perspective of the color and shape features,the content-based image retrieval technology is applied to pedestrian surveillance video retrieval in buildings,pedestrian image retrieval experiments are designed and implemented based HSV color space and SIFT feature.The results show that the above method is not good atpedestrian image recognition and low retrieval accuracy.Then the experiment of feature extraction and retrieval based on ResNet50 network,a pre-trained network of Keras framework,is designed.The results show that the retrieval accuracy has been improved,but the computation is heavy and time-consuming.According to the above experimental results,a two-stage retrieval algorithm based on HSV color space and ResNet50 network is proposed,and then the design of pedestrian surveillance video retrieval system in buildings is completed using that algorithm.The video retrieval system is constructed using tensorflow,a open source framework,and the architecture design,video processing,pedestrian detection,pedestrian retrieval and video cutting of the system are described in detail.Finally,PyQt5 is used to complete the application program of pedestrian surveillance video retrieval system in buildings,so as to realize intelligent pedestrian video retrieval and improve work efficiency.