Research On Infrared Small Target Tracking Technology Based On Deep Neural Network

Compared with visible light imaging,infrared imaging system has the advantages of longer operating distance,stronger anti-interference ability and all-weather operation.It is widely used in the fields of early warning system,air defense system,precision guided weapons and so on.Infrared small target tracking is the key technology and has important research significance.In order to track infrared targets more accurately and efficiently,this paper studies the infrared small target tracking technology based on deep neural network,and the following works have been achieved.(1)Aiming at the poor tracking performance of infrared small targets in complex environment,an infrared small target tracking algorithm based on attention siamese network is proposed.Channel attention module based on multi-scale features and feature fusion structure are applied together to improve the distinguishability of target and background,and a training strategy based on time effectiveness is proposed to improve training performance.Experimental results show that Inf Siam Net effectively improves the tracking performance,especially the stability in complex environment.(2)Aiming at the tracking difficulty caused by target maneuvering,an infrared small target tracking algorithm based on transformer is proposed.The algorithm makes full use of the superior ability of transformer in capturing long-distance dependence and parallel computing,extracts the spatial and temporal features of infrared images,and makes up for the less available features and low recognition of infrared targets.The algorithm regards tracking as a parallel target box prediction problem and simplifies tracking pipeline.Experiments show that the tracker significantly improves success rate and accuracy,especially when the target maneuvers.(3)In order to improve the robustness of tracking,an integrative processing algorithm of detection with tracking based on SSD and Inf Siam Net is proposed.The lightweight MobileNet network and feature fusion structure are applied to improve SSD detection,and the detection and tracking work together.Experimental results show that the tracking performance of this algorithm is more stable compared with the single target tracking algorithm.