Chinese Journal of Computers   Full Text
Title	The Detection of Multiple Targets with Low SNR Based on Greedy Algorithm
Authors	ZHANG Hai-Ying1) WEN Xuan2) ZHANG Tian-Wen3)
Address	1)(Software School of Xiamen University, Xiamen University, Xiamen, Fujian 361005) 2)(Xiamen University Tan Kah Kee College, Xiamen University, Xiamen, Fujian 363105) 3)(School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001)
Year	2008
Issue	No.1(142—150)
Abstract &   Background	Abstract Aiming at the detection of dim targets with SNR≤2dB, Track before detection method based on greedy principle is presented. The state update is completed by choosing an extended node and the energy is accumulated along the trajectory. Finally the target is identified according to likelihood ratio test. In pre-processing stage, spatial-temporal filter is designed. On one hand, clutter is de-correlation based on compound kernel estimation spatial filter; On the other hand, noise is restrained via projection along time axis, the searching space is reduced about 80%. Furthermore, the independence and normality of residuals after kernel estimation are tested in nonparametric statistical theory. The comparison experiments with traditional dynamical programming algorithm show that the time complexity can reduce to mn. It not only has the predominant in time but also has higher probability of detection and lower false alarm than the former. keywords kernel estimation; point targets; extended node; time-projection; greedy algorithm background This work is supported the by Fund Supporting Aerospace Technique（HT010415） and the National Natural Science Foundation of China (60672018).The subject of the paper is in the field of ATR (Automatic Target Recognition). The main purpose is to research the problem of detection and tracking of dim point targets in complexity background. When a target of small size(<10m in length) is remote from the sensor(>100km), it is imaged at only one pixel or less than one pixel. This type of target is referred to as "pixel-sized" target or "point target". The difficulties of the detection and tracking task are clear: For pixel-sized targets, conventional pattern recognition methods fail for lack of shape information; some randomly distributed high-intensity noise pixels have the same appearance as the targets in a frame. With little knowledge about the trajectories in the time sequence, the task becomes extremely difficult. Consequently the topic draws much attention of researchers national and international. The similar research work has been carried out about several decades abroad and many classical methods and theories have been presented, such as MHT, DPA, NN and wavelet. But they all have shortcomings, especially the resolution of key problem is relying on the apriority information of targets dynamical equation and that is difficult to obtain in many circumstances. In this paper, the data association is accomplished by the highly relativity of targets in time and space without these information. The validity has been proved by experiments.