Design And Implementation Of Pedestrian Detection System In Intelligent Video Surveillance

With the development of society,more and more video surveillance devices are used for public security,and pedestrian detection is an essential part of ensuring public safety.However,the current pedestrian detection technology has obvious omissions and false detections for pedestrians in practical applications,and is easily affected by environmental changes,pedestrian attitudes,pedestrian scale and other factors,resulting in insufficient robustness and poor detection results.In order to improve the pedestrian detection ability in the video surveillance scenes,this paper fully considers the various influencing factors in the actual monitoring scenes,and proposes a residual learning network YOLO-CRD that combines shallow visual features and deep semantic features.Based on this,a multi-task loss function optimized for adaptive pedestrian scale is designed to optimize YOLO-CRD.The main research work and results of this paper are as follows.(1)This paper uses traditional pedestrian detection algorithm and machine learning based pedestrian detection algorithm to detect pedestrians in video surveillance,and analyzes and summarizes the principle and existing problems of each method.(2)For the problem that YOLO has low detection accuracy for pedestrians with small scale,this paper improves and proposes a pedestrian detection network YOLO-CRD with enhanced small pedestrian detection that combines shallow visual features and deep semantic features.And additionally the proposed method is applied to design and implement the pedestrian detection system to detect pedestrians in video surveillance.(3)In order to further strengthen the detection algorithm's ability to detect pedestrians,this paper designs and optimizes an adaptive pedestrian-scale multi-task loss function,and augments the training images containing pedestrians according to the characteristics of pedestrians' attitude changes and scale changes in video surveillance.Corresponding image data augmentation is performed to enhance the pedestrian detection performance in different forms.(4)Using the pedestrian detection system in video surveillance,we can achieve a)to build a client to provide an algorithmic visualization interface;b)to use the improved YOLO-CRD algorithm to detect pedestrians.c)to automatically save the images of the pedestrian detected for later retrieval.Experiments show that even in complex scenarios,the accuracy and real-time performance of the system meets the requirements and has good performance.