Research On Knowledge-Based Target Recognition And Tracking Techniques

With the development of science and technology, automatic target recognition and tracking techniques have become a very active topic in image processing and computer vision field. It plays an increasingly important role in civilian and military applications. Invariant features have good numerical stability for the same targets in different status, so it becomes a crucial step in target recognition. Artificial neural network provides a new approach for automatic target recognition and tracking, which makes the process more intelligent. The rapid development of hardware recent years, makes it possible to develop real-time tracking systems. Considering the limitation of the traditional statistical pattern recognition, a recognition method based on moment invariants and artificial neural network is proposed in this dissertation, and also a hardware system is designed.The main work of this dissertation is listed as following:First, the extraction of invariant features is studied. We choose moment invariants as the feature in this dissertation, which has good stability for geometric transformation. The basic concepts of moments are expounded, including the definition of geometric moments and geometric transformations. Various moment invariants are studied, including the traditional moment invariants which are invariant to continuous image, the discrete moment invariants which are invariant to discrete image, and the contour moment invariants which are invariant to contour image. The invariance of Hu moments is proved, and on the basis of the proof, we represent a formula to construct new moment invariants. 5 new moment invariants are constructed using the formula, and experiments show that the 5 new ones are as invariant as the original ones. The performance of regional and contour moments is also compared.Then, BP neural network is researched, including network structure and learning process, etc. Against its limitations, a number of useful improvements are proposed in order to enhance its performance. Feature extraction and feature selection is discussed, then some effective features are chosen from all the moment invariants. Using these effective moment invariants as the feature of the target, and BP neural network as the classifier, some experiments on the recognition of several targets are implemented. Experiments show that the method can achieve a good recognition rate.Finally, on the basis of the above theory, a tracking system is designed and implemented. The system consists of two subsystems: the image input-output subsystem and image processing subsystem. The latter subsystem is designed using a DSP + FPGA architecture. To solve the low efficiency of the system, we also do a lot of work on the optimization. Experiments show that the tracking system has excellent reliability and real-time property.