A Study Of Object Tracking Based On Local Model And Biomimetic Pattern Recognition

Object tracking plays an important role in the field of computer vision, with the goal of locating the target and capturing the change of target’s appearance and motion in each frame. It has been widespread concern all around the world, and has been applied to intelligent monitoring, medical imaging, car navigation and other projects. This paper summarizes classic theories and algorithms in the field of object tracking, and gives comprehensive analysis on the technical points and the evaluation criteria. Biomimetic pattern recognition is a novel pattern recognition the idea of "cognition" instead of the traditional idea of "division", its implementation is multi-weight neural network. Based on these theories, the author’s major contributions are outlined as follows:1. The training samples of object tracking are obtained according to the sampling time, sometimes, there are pseudo targets in the background which would interfere with acquiring the real target. So this paper proposes an improved method of multi-weight neural network, which is based on distance. Considering the relationship between the target of the current frame and the training set, we introduce the distance weight. Experimental results are presented to demonstrate that this method can effectively reduce the pseudo targets interference.2. In recent years, the local model has been combined with some classical tracking methods for handling occlusion, shape and illumination changes. Combining the local model with biomimetic pattern recognition method, this paper proposes the patch biomimetic pattern recognition based object tracking method. We cut training samples into patches, and obtain a set of patch training samples which are chosen to train the neural network. The candidate regions are put into these neural network, each set will get its closest results, from which we can locate the target. Experimental results are presented to demonstrate that this method can be adaptive to occlusion.3. Image distance metric plays an important role in image processing, pattern recognition, and many other fields. To identify the tracking target, we need to calculate the distance between candidate samples and topologically set which is constructed by multi-weight neural network. Generally, the Euclidean distance is used to do this work, however, it does not encode any discriminative information. This paper proposes the local standard Euclidean distance, which process the local image within a radius. Experimental results are presented to demonstrate that the metric can be adaptive to occlusion.