Research On Dynamic Object Recognition And Tracking Algorithm Based On Deep Learning

Vision-based environmental perception technology has been applied to various aspects of daily life.Therefore,it has important research value as well as practical significance.With the complexity of application scenarios,environmental perception capabilities such as object recognition and tracking should satisfy higher requirements accordingly.However,the accuracy and speed of existing object recognition and tracking technology in indoor occlusion and outdoor open scenarios cannot satisfy the actual demands,and cannot perform the task of environmental perception excellently.This paper uses deep convolutional network technology to study object recognition and tracking algorithms of dynamic armed objects such as armed personnel and firearms to improve the effectiveness and robustness of visual perception algorithms in practical scenarios.The main research contents are as follows:First,for armed object datasets such as Pistols and OP-Weapons,there are problems such as simple environment,single perspective and type,an armed object dataset containing complex environment and multi-perspective features is constructed.Aiming at the problem of missed detection and false detection caused by similar feature interference,an armed object recognition algorithm combining fine-grained feature extraction method and feature pyramid technology is studied.The Darknet-53 network with residual structure is applied to achieve refined feature extraction,and based on the feature pyramid idea,a multi-layer feature fusion strategy is designed to realize accurate recognition of objects such as armed men and firearms under a single camera.Second,aiming at the problem of limited field of view of the single camera,a multi-camera multi-target tracking algorithm based on space-time constraints is designed.Aiming at the problem of false detection and missed detection caused by long-term occlusion between dynamic targets,the task of matching the target trajectory is completed through the complex state space matching mechanism and timing information auxiliary constraints.For the impact of similar feature interference on the accuracy of data association,a data association method combining spatial position constraints and hierarchical clustering algorithm is studied,which further constrains the single-camera trajectory to improve the data association accuracy and complete the global perception of armed objects.The effectiveness of the algorithm is verified by comparing the experimental results with its peers on public datasets.Finally,the armed object recognition and tracking algorithm is testified in different actual scenarios to verify the effectiveness and generalizability of the algorithm.