Research On Infrared Target Detection And Tracking Algorithms Under Complex Background

Infrared imaging technology works by receiving targets'infrared radiation passively. It has such advantages as long operating range, convenience for hiding, and ability to work double tides. With the continuous development of infrared imaging technology, infrared imaging system has been applied to many military or civil fields, such as infrared precise guidance, early warning, video surveillance, search and tracking. As a key technique in the above fields, target detection and tracking based on infrared imaging play important roles in modern defense. Especially, the problem of infrared target detection and tracking under complex background has been the research focus in recent years. Under complex background, infrared targets have their own characteristics which make the detection and tracking work become quite difficulty. For example, for imaging distances of targets are always far, they only occupy several pixels in the image. In addition, both of noises in the imaging system and background clutter disturbances are so strong that the target signals are dim and easily drowned by them. Finally, the targets lack of efficient shape and texture features, so little information can be provided for the detection and tracking system. Therefore, the detection and tracking of infrared target under complex background are challenging research topics, and an in-depth study of them has crucial theory meaning and practical values.This dissertation deeply discusses and researches the related technology about the detection and tracking of infrared target under complex background. The innovative contributions of this dissertation are as follows:(1) An algorithm based on wavelet packets and higher-order statistics is presented for infrared dim and small target detection. To overcome the defects of traditional wavelet transform based target detection method which is vulnerable by background clutters and noise disturbances, the characteristics of infrared dim and small target images are firstly analyzed. On the basis of it, the wavelet packets are firstly used to decompose the image into multiple scales, which can overcome the limitation of the higher frequencies having lower resolutions in wavelet transform. Then, to eliminate the influence of noises better, a Gaussian measurement criterion based on higher-order statistics is proposed, which can select the corresponding frequencies to the frequency spectrum of the target adaptively. Finally, the satisfying detection results are achieved. Since the proposed algorithm combines the advantages of wavelet packets in signal decomposition and those of higher-order statistics in signal analysis, it not only effectively raises the target detection probability, but also reduces the false alarm probability compared with the classical wavelet transform based target detection method.(2) Two improved methods based on the fractal theory are proposed for fast infrared dim and small target detection under complex background. The first improved algorithm is the local entropy based fractal method. Local entropy reflects the discreteness of the image gray. The gray distribution of the region which has higher local entropy values is homogeneous, while the gray distribution of the region which has lower local entropy values is heterogeneous. So the presented method uses local entropy of the image to locate the target coarsely. Then a fractal dimension image based target detection method is proposed to detect the infrared dim and small targets precisely.The second improved algorithm is the third-order characterization based fractal method. For the infrared dim and small targets are always thought as transient non-Gaussian signals, the local region where the target exists in deviates further from normal distribution than other regions in the image. Therefore, the presented algorithm firstly uses the third-order characterization to segment a region of interest which contains the target, and on the basis of it locates the target finely by fractal theroy. Compared with the traditional fractal algorithm, each of these two improved methods can be divided into two parts:coarse location and accurate location. The coarse results not only improve the robustness of dim and small target detection process effectively, but also reduce the region to be processed by the fractal algorithm to a small range, which raises the efficiency of the method greatly.(3) Two improved Particle Filter methods are proposed for infrared target tracking. Firstly, to solve the problem of infrared dim and small target tracking under complex background, an improved Particle Filter algorithm is presented. For infrared dim and small target, gray and fractal are two very important features. Each of them has advantages as well as limitations:the tracking only using gray feature usually has deviations or errors, while fractal feature is only used for infrared dim small target detection problems in a single image currently. To make the most of these two features, this dissertation fuses them together in the probabilistic model of Particle Filter for the first time, and uses the fusion results to calculate the particles'weights, which greatly improves the robustness and accuracy of the tracking algorithm.In addition, based on the in-depth study of infrared human target tracking under complex background, an improved Particle Filter algorithm based on the gray and motion cues is proposed. The algorithm firstly constructs the gray histogram which possesses the spatial localization information of the target to extract the gray feature, and at the same time proposes a new method based on the inter-frame difference and gray probability distribution map to get the motion feature. Then, a probability fusion of these two features is performed in the Particle Filter framework. Finally the robust infrared human target tracking is completed. Compared with the traditional Particle Filter algorithm, the presented method can handle strong background disturbances as well as occlusions much better, and has higher tracking precision and stronger stability.(4) This dissertation extends the Mean Shift tracking framework from two points of view: the adaptive fusion of multiple cues and the integration with Particle Filter framework. Starting from the former angle, an improved Mean Shift algorithm based on adaptive multi-cue fusion is presented for infrared human target tracking. It extracts the gray and edge features of the target at first, and then uses the motion information to guide the two features to obtain the new target cues:motion-guided gray cue and motion-guided edge cue. Subsequently, a novel adaptive fusion scheme is proposed to integrate these two novel cues adaptively into the Mean Shift framework. At last, an automatic target model update strategy is proposed to improve the tracking performance further. The improved method can handle many complex situations, such as background clutters, changing illumination, changing appearance of the target, as well as occlusions. And the tracking results are satisfying.According to the latter perspective, an algorithm which integrates the advantages of Mean Shift and Particle Filter frameworks is proposed for infrared target tracking. The algorithm firstly improves the traditional target model and presents a novel target model, which fuses the color and motion cues, to enhance the robustness and accuracy of target tracking. Then it proposes an improved Mean Shift method, and then embeds it into the Particle Filter framework to rearrange the random particles, in which the particles move toward the maximal posterior probability density of the target state. Thus, the efficiency is raised. Experimental results show that the proposed method is significantly superior to the traditional Mean Shift algorithm or Particle Filter algorithm, and it can also guarantee fast and precise target tracking under the situations such as the fast moving targets and occlusions.