Studies On The Algorithm Of Target Segmentation And Tracking Based On Wide-field-of-view Image Mosaic

The exploration of mineral resources in the field of civil, land use planning,environmental supervision,ocean developing industry,weather forecast and geographical information service,or investigation and surveillance in the military field,precision guidance,over the horizon offensive and defensive confrontation etc. All of them need to have a wide field of view and high resolution to complete the target of a large range of monitoring, search and tracking. For high resolution imaging system based on wide field of view, massive image data processing, analysis and utilization is the core value of the system construction. Among them, the high resolution sensor mosaic imaging process will involve massive data real-time processing, computational complexity of the process is very high.How to ensure the accuracy and real-time of image mosaic algorithm is one of the key problems to be solved in the development of the system. In addition, the dynamic object tracking technique with wide field of view and high resolution scene has become a research focus of image analysis in the later stage. At the same time, due to the complexity of the environment of the system application(background change, illumination change, shadow change, etc.)and some properties inherent to dynamic targets(non rigid body, attitude change, etc.),it is still a challenging problem to be solved in the practical application of target tracking technology. In view of the above requirements, the problem of image mosaic and object tracking in wide field imaging system is studied in this paper. The main research work is as follows:In order to realize fast imaging of large field of view panoramic mosaic,an adaptive image mosaic parallel algorithm based on unified computing device architecture(CUDA) and a priori information is proposed. Firstly, high precision calibration platform is used to pre calibrate the overlapped area of the adjacent micro camera. Then, a fast robust feature detection method based on CUDA is used to detect the candidate feature points in the overlapped area. Then, an approximate nearest neighbor search algorithm based on the random KD-tree index is accelerated by using the basic linear algebra operations,used to obtain the initial matching point pair. Finally, an improved parallel incremental sampling consistency algorithm is proposed to eliminate the error matching point pair and parameter estimation of spatial transformation matrix. Accordingly the spatial geometric transformation of the image is obtained. Experimental results show that compared with CPU serial algorithm, the algorithm speed used in this paper increased and it can meet the requirement of real time for image mosaic in engineering application.In order to identify the dynamic flight targets an image segmentation method based on chaotic double population evolution strategy is proposed. The evolutionary strategy can be used to set out from the initial solution,through continuous iteration to improve the current solution,until the end of the search to the optimal solution or satisfactory solution of the characteristics and advantages, It is used to solve the optimal solution of image segmentation threshold. In order to overcome the shortcomings of the traditional image segmentation method based on threshold,for example, the high complexity and premature problem,an efficient algorithm for image segmentation based on evolutionary strategy is proposed in this paper. It uses a variety of cluster evolution strategies to calculate thresholds. There are two populations in the process of evolution- local population and global population, and then ensure the global and local search capability of the algorithm. In each iteration step of the algorithm, Firstly, based on the chaos theory, a number of initial individuals are generated. These individuals were added to the local population and the global population. Calculate the fitness function value of these individuals. Then, select the local population and the global population in parallel, recombination, mutation and other evolutionary operations. In the set of evolutionary individuals, select some of the best individuals into the local population, put the rest into the global population until convergence. Finally, the optimal individual in the population is the solution. Experimental results show that the method proposed in this paper is better than the traditional method based on genetic algorithm.Population diversity information can effectively guide the evolution process of evolutionary strategy. Therefore, this paper proposes an improved chaos double population evolution strategy algorithm, the multiple motivation reinforcement learning algorithm is used to set the initial population and the local population, which can learn population ratio dynamically in order to further balance the evolutionary strategy ability of local search and global exploration. The introduction of the motivation layer provides the conditions for the introduction of prior knowledge and domain knowledge. This can accelerate the learning process of reinforcement learning. In this paper, according to the image segmentation problem, defines a set of motives, use MMQ-voting method to guide the selection strategy of intelligent body movements. After testing and verification, the reinforcement learning method based on multi motivation can make the reinforcement learning to converge to the optimal action strategy at a faster speed, to keep the diversity of the population in a suitable state. It is helpful to further improve the search efficiency of the image segmentation threshold.In order to track the dynamic flight targets a dynamic target tracking method based on reinforcement learning is proposed. The target tracking problem is modeled as a reinforcement learning problem. A two stage reinforcement learning algorithm is proposed for the target tracking in the image. We set up multiple tracking agents to track the target in the image. In every step of the algorithm, firstly, the dynamic sub task allocation for each tracking agent is carried out, that is to give each tracking agent dynamic allocation of a sub target. Afterwards each trace agent then selects its actions based on its current child target. The learning algorithm divides the learning process into two parts, One is the learning task allocation strategy, another is the choice of learning strategies. Each trace agent shares the learned knowledge by sharing the Q function and improves learning efficiency. The experimental results verify the effectiveness of the proposed method.