Infrared Small Moving Target Detection And Tracking And DSP Real-Time Implementation

Infrared small moving targets is widely used in disaster relief,video monitoring and military reconnaissance.However,in these applications,target detection and tracking face challenges such as long detection distance,target with small image area,complex background,and the interference from moving platform.Moreover,these applications also have high requirements for the real-time performance,small volume and low power consumption.Therefore,it is of great significance to study real-time detection and tracking of small infrared moving targets based on the embedded platform.This thesis focuses on the following specific issues in the above applications.The fixed-size filter is difficult to adapt to the multi-scale targets in the image,meanwhile multi-scale detection and fusion will bring about the problem of large computation and high false alarm rate.To address these issues,we study the approximate model of target gray distribution,and an infrared dim small targets scale estimate algorithm is proposed.The algorithm estimates the possible size of the target pixel by pixel.Then,the size of the filter on different locations is determined according to the estimated target size.Experimental results show that the proposed adaptive scale detection algorithm can double the SNR gain and increase the speed by 2 to 3 times compared with the multi-scale detection.In the literature of keypoints registration based moving targets detection,sometimes the target is so complex that some registrated keypoints might be densely located on the target,which leads to the problem that platform motion estimation being disturbed by target motion and the target would be subtracted after frame difference as the result.To solve it,this thesis proposes a keypoints redistribution algorithm.The algorithm deletes some keypoints that are too dense in a local image patch so that they would be more evenly distributed in the whole image.Experiments show that the proposed algorithm can obtain platform motion estimation and frame difference results more accurately with higher signal-to-noise ratio without affecting the registration accuracy,making the target easier to be segmented.In the field of correlation filtering tracking,the model updating strategy with fixed learning rate is difficult to deal with the situation that the target changes rapidly or slowly.For the problem,a new correlation surface quality evaluation metric is proposed in this thesis.Meanwhile,we consider the occlusion of the target,and propose the occlusion judgment algorithm.Finally,the learning rate of model is updated adaptively according to the change speed of the target and whether the target is occluded.We further conduct experiment on the OTB50 dataset.Experimental reults show that the correlation tracker with gray feature can improve the tracking accuracy by about 3%,after adopting the learning rate adaptive strategy proposed in this thesis.Better tracking result is also obtained on the infrared sequence that photographed in real scenes.Aiming at the real time performance and small volume of system,the algorithm is optimized for DSP platform.The precision error between the implementation result and the simulation result is analyzed,the running time on DSP platform is tested.Experiments show that the proposed algorithm and DSP implementation scheme in this thesis meet the requirements of the reserch in terms of accuracy and real-time performance.