Research On Fast Obcjet Finding Technology In Video

The fast search of the object in the video aims to quickly find the segment of the interested Object search aims to find and match people and cars from surveillance videos.It is an important link in the construction of intelligent surveillance systems and smart cities.It has received extensive attention and research from academia and engineering circles for its important use value.The mainstream object finding method is based on object detection and object retry.Among them,it is difficult for object detection to meet the requirements of speed and accuracy at the same time,high-efficiency detectors are easy to miss objects,and highprecision detectors cannot achieve real-time calculations.The object re-recognition still has a large distance between research and application,and the matching speed also has room for improvement.This article mainly focuses on improving the detection and re-recognition speed under the premise of ensuring the accuracy of the object detection network and object re-recognition,carrying out theoretical research and extensive experimental verification,and combining the research results to achieve a object search system in the video,which satisfies the perception of the video.Quick search of interest objects.The main work and innovations of this paper are as follows:1.On the object detection network based on Yolov5,Ghost Net and SE modules are added,which reduces the amount of network parameters and improves the processing speed of the network.Experiments show that the Yolov5 network with Ghost Net and SE modules has a reduced amount of parameters and an increased detection speed while the accuracy is basically the same,achieving faster target detection.And based on the improved Yolov5 target detection network,a target detection network for people and vehicles is trained on the pedestrian detection data set and vehicle detection data set created by combining multiple datasets.2.Improve the object re-identification network in terms of optimized matching structure and model compression.In terms of optimizing the matching structure,a multi-dimensional feature matching method is proposed in which feature vectors of multiple dimensions are matched from low to high under a set threshold.Multiple dimensional feature vectors are generated by a feature extraction network through similarity knowledge distillation,which avoids the need to use multiple feature extraction networks to generate multiple dimensional feature vectors,and at the same time strengthens the accuracy of low-dimensional features;the threshold is passed through Threshold optimization algorithm to calculate,compared to exhaustively to find the optimal threshold shortened the time.In terms of model compression,replace the backbone of the network with Res Net18,and use the method based on soft label knowledge distillation to "transmit" the knowledge of the trained backbone as the Res Net50 network to the new backbone as the Res Net18 network.Combining these two improvements,the inference speed of the object re-recognition network has been accelerated by 3 times,and the matching speed has been accelerated by 15 times.3.Combining the search algorithm designed with object detection and object reidentification and the easy-to-operate interface made by Tinker,a system for quickly finding objects in the video is realized.The processing flow is to first use the trained object detection network for people and cars to detect the object in the video frame,and then use the trained person or car re-recognition network for the detected object to match the search object,and the final output Find the video segment where the object is located and mark the object in the video segment.Experiments show that the system completes the search for a specific object in the video within one-tenth of the time to find the video,which is of great significance to the security field.