The Research On Dim Small Target's Detection In Infrared Image Sequences

Nowadays, as high-tech weapon technology develops rapidly in the world, infrared detection system has become one of the key researching and developing projects all over the world due to its good concealment, strong anti-jamming performance and high tracking accuracy, and it is widely used in various weapon equipments. Recently, as the requirements in the military fields are increasing, study on the moving target detection technology under far distance and low-SNR (signal-to-noise ratio) conditions becomes a hotspot. This thesis studies the infrared dim small target detection problem in depth, and mainly includes detection-before-tracking based on wavelet analysis algorithm and tracking-before-detection based on particle filtering algorithm.First, according to the spatial math model containing the infrared dim small target, by using gray singularity feature of the dim small target, a kind of wavelet base function is selected for the dim small target detection, thus presenting multi-scale fusion detection based on wavelet analysis method. Then according to the target motion continuity and the track consistency in the sequential images, and the random property of the false target and noise motion, the dim small target is identified by using pipeline filtering. Simulation results show that this algorithm can efficiently detect the infrared dim small moving target with SNR above 2.5, with less memory and computation complexity.For the second algorithm, characteristics of the dim small target's temporal model are first analyzed, and then an infrared dim small target detection algorithm based on 1-D and two-level wavelet transform is presented. It decomposes the original image sequences into sub-images with different feature and different frequencies, and then processes the high-frequency sub-images, thus detecting the target. After that, by combining the morphologic spatial filtering with this algorithm, the temporal and spatial fusion detection algorithm is realized, and it overall considers the small moving target characteristics both in spatial and temporal domain and gets over the limitation of target detection just from one side.Simulations for the two detection algorithms above are done, and test results show that for the detection-before-tracking algorithm, when the SNR is below 2.5, the detection probability is not high while the false probability is higher. Thus, when the infrared target is too small to be detected efficiently, some tracking methods could be used before detection to estimate the target position in the spatial plane, and then some detection algorithms are used to detect and identify the estimated track so as to achieve small target accumulation effect along the track to improve the detection performance, and that is just the essence of the tracking-before-detection algorithm.Tracking-before-detection based on particle filtering algorithm is the study hotspot in recent years.Aiming at the degeneracy problem in the PF (Particle Filtering) algorithm, several resampling algorithms and choice of the importance function are studied in depth. For nonlinear/non-Gaussian conditions, the tracking performance of PF is superior to that of EKF (Extended Kalman filtering). For strong nonlinear/non-Gaussian systems and high precision conditions, tracking performances of EKF and PF is not good, and via simulations for PF, UPF (Unscented Particle Filtering),) and UKF (Unscented Kalman Filtering), results show that tracking precision of UPF is obviously higher than that of PF and UKF. But UPF has a problem that its time costs are higher than that of PF and UKF. Therefore, this thesis puts forwards an improved UPF algorithm. The tracking precision of the improved UPF algorithm is equivalent to that of UPF, but with 35% time costs of the latter.Aiming at the moving target detection under low SNR, realization steps of the tracking-before-detection (TBD) based on particle filtering algorithm in detail is studied, also improvement on it is put forward. Threshold for sampling particles at the prediction stage is set, and it not only improves the validity but also reduces computational complexity. Via simulations, tracking performances, detection performances and effects of the two algorithms by the particle number under different SNR are compared, and conclusions are given. By applying tracking-before-detection based on particle filtering algorithm in the infrared dim small target detection, and according to the infrared image characteristics, a tracking-before-detection scheme for the whole view field is presented. Simulations for the image sequences with different SNR are done, and detection and tracking performances of this scheme are obviously superior to that of the detection-before-tracking algorithm, but with higher time costs.