Research And Implementation Of Infrared Small Target Tracking Algorithm

Target tracking is a very important topic in the field of computer vision.With the development of science and technology,target tracking has developed rapidly and is widely used in military industry and civil life.However,due to factors such as small target scale and complex background,there are still great challenges in tracking accuracy and tracking speed.This paper focuses on infrared small target tracking and studies in the context of air combat investigation.The image often has a very small target because of the long shooting distance and the influence of noise and cloud occlusion.Even the target has only 1-2 pixels.Due to the low SNR and the small number of target features that can be extracted of such a very small infrared target,most tracking algorithms cannot perform normal tracking at present.Moreover,the tracking algorithm with good tracking performance is relatively computationally complex.As a result,the tracking speed is slow and cannot meet the requirements of high-speed tracking.Therefore,improving the speed of the tracking algorithm is also an urgent problem to be solved.KCF(Kernelized Correlation Filters)is a fast tracking algorithm that detects the next frame by training the classifier between the target and the background.However,its accuracy and speed in the tracking of infrared small targets still cannot meet the demand.In this paper,a KCF_MFD tracking algorithm based on KCF algorithm is proposed for very small infrared target tracking.The KCF_MFD tracking algorithm uses multi-scale flux density as the feature of small targets for tracking.The tracking accuracy and bounding box overlap rate are robust in small infrared target scene tracking.Because the FPGA has the characteristics of parallelism and fast calculation speed,this paper uses FPGA to implement the KCF tracking algorithm,and finally realize the high-speed tracking system.This paper will focus on the research of very small infrared target tracking and FPGA implementation of KCF algorithm.The main research contents and research results are as follows:Add the multi-scale flux density of the small target as the target feature to the KCF framework for tracking to obtain the tracking algorithm KCF_MFD.It is tested by OTB,and it has good performance in the center position error and bounding box overlap rate in the tracking of very small targets.The tracking accuracy score is 0.886 and the success ratio score is 0.439.The high-speed tracking system is implemented by FPGA implementation of parallel KCF algorithm.The system mainly comprises a DDR cache module,a control module,an image feature extraction module,a training module and a detection module,which can realize high-speed tracking.The system is implemented in the Virtex-6 XC6VLX240T-1FFG1156 model FPGA,and the resource usage slice is 46%.Using the VOT2014 dataset to test the system.The maximum delay is less than 3ms and the average delay is 1.676 ms.The minimum throughput is more than 800 FPS,and the average throughput is 1466 FPS.The system has a good performance in terms of resource occupancy and tracking speed.