Research On Infrared Security Multi-target Tracking Algorithm Based On Lightweight Models

With the application and promotion of infrared imaging technology,more and more infrared technology is used in the field of intelligent security and transportation.In the field of security,monitoring of the surrounding environment by infrared cameras can achieve 24-hour comprehensive monitoring.Due to the low contrast and less detail of infrared images,making it lack of more information compared to ordinary visible images,increasing the difficulty of target location tracking on infrared images,more accurate identification and tracking of people and vehicles in infrared images is one of the current technical challenges of intelligent security.Most of the current multi-target tracking algorithms for IR target tracking have problems of high computational complexity,low accuracy and poor real-time,which make them difficult to deploy on removable edge hardware platform devices and cannot meet the practical application requirements.In this paper,we address the above problems by investigating infrared security multi-target tracking algorithms based on lightweight models,with the following main work.(1)For the task of real-time detection of infrared targets,the detection capability of the deep learning model YOLOv5 for infrared targets is studied.In order to reduce the number of parameters of the model as well as to improve the computational speed of the model during detection,making it more suitable for security surveillance devices,this paper uses the Ghostnet module to optimize the structure of the YOLOv5 s model,and the improved obtained GhostYOLOv5 s In addition,this paper also uses MCSP data enhancement and FSP matrixbased knowledge distillation to improve the target detection capability of the improved model,especially for small infrared targets with pixel areas below 94.8% of the mAP obtained from the improved model in the test set.(2)For the real-time infrared target tracking task,the tracking performance of the real-time target tracking algorithm DeepSort algorithm for infrared human-vehicle targets is studied.In order to improve the overall operation speed of the algorithm,the OSNet model is used to extract the apparent features of multiple types of targets in the DeepSort algorithm,which reduces the total data volume of the algorithm by 94.3% and increases the computing speed by nearly four times.In order to alleviate the phenomenon of frequent switching of target IDs in the multi-target tracking process,the calculation method of IOU weights in the second stage matching in DeepSort is improved,and the multiple difference characteristics between the detection frame and the prediction frame are introduced to make the weight matrix values more discrete,so as to effectively reduce the switching frequency of target IDs in the tracking process.(3)The improved IR multi-target tracking algorithm is deployed on the Jetson Nano hardware platform,and the MOTA in the test set is 76.7,and the algorithm runs at 68 FPS in the hardware platform.The experimental results show that the improved algorithm has the slowest performance degradation compared with other multi-target tracking algorithms,indicating that the improved multi-target tracking algorithm has good anti-noise performance.Finally,the intelligent infrared surveillance system is designed and implemented based on the hardware platform and the improved multi-target tracking algorithm,and the actual tracking effect in different infrared scenes is tested,which achieves good security monitoring requirements.