Research On Infrared Small Target Detection Technology Based On Trajectory Feature Learning

Infrared target detection technology plays an increasingly important role in satellite remote sensing mapping,photoelectric detection,and space situational awareness.Among them,the problem of infrared small target detection is the technical difficulty and task focus.Infrared small target refers to the target which appears as light spot after being imaged by the photoelectric sensor due to the long imaging distance or the relatively small shape of the target itself.Compared with the traditional detection target,it lacks the information of shape and texture.This paper focuses on the trajectory detection of small infrared targets,comprehensively utilizes the spatiotemporal information,and associates the preprocessed time series images of candidate points as target trajectories to further improve the detection performance of small infrared targets.(1)Analyze and sort out the process of infrared small target trajectory detection,summarize and implement three single-frame detection methods and analyze their detection performance.Based on satellite remote sensing data and target motion model,the infrared small target motion dataset of uniform motion,accelerated motion and turning motion with acceleration disturbance was constructed.(2)A multi-hypothesis testing method based on intelligent scoring of trajectory anomaly detection is proposed.A multi-hypothesis gate correlation module is designed for the low signal-to-noise ratio infrared small target with many candidate noise points and the target trajectory state is easy to lose.Aiming at the dependence of traditional multi-hypothesis testing algorithms on prior parameters,a trajectory intelligent scoring module based on "one-class" anomaly detection is designed.Compared with the "twoclass" method,it avoids the process of simulating abnormal trajectory data sets.The candidate trajectories are subjected to the reconstruction error threshold detection based on statistical distribution to judge the authenticity of the trajectory.Compared with the traditional multi-hypothesis testing method,the detection performance is improved,and the intelligence level and iterability are improved.(3)An infrared small target detection algorithm based on point cloud feature learning is proposed.In this paper,the spatiotemporal information of the target is fused,and the concept of an alternative point cloud is proposed.Aiming at the non-uniformity of the target trajectory point cloud,a multi-scale clustering module is designed to retain the multi-scale features of the target trajectory point cloud.According to the similarity of the topological structure of the target trajectory and the fixed background noise in the candidate point cloud space,a targeted loss function is constructed to further reduce the false alarm rate.The effectiveness of the network for a variety of trajectory scenarios is verified,and the detection performance of the target trajectory point cloud is improved compared with the traditional point cloud target detection network.