Research On Visual Tracking And Action Recognition Based On Bag-of-Words Model

Visual target tracking and action recognition are hot research problems in comput er vision and pattern recognition. Due to the variations of visual angle, posture, size, li ght and occlusion problem, it is still difficult to track target with robustness in long ti me. Besides the variation of external environment, there are still some other issues, su ch as no unified form with same action and difficulties in action segmentations in time dimension, which makes action recognition a challenge problem. Based on Bag-of-W ords model, two algorithms are proposed to solve the problems of target tracking and action recognition. One is a novel tracking method based on bag of features and partic le filter with dense sampling, and the other is a novel action recognition method based on the space time points and bag of features.To begin with, we propose a novel tracking method based on bag of features and particle filter with dense sampling. Firstly, local image patches within a object region are densely extracted by using sliding windows (DSIFT) and represented as invariant descriptors. Secondly, visual codebooks are generated by clustering algorithms such as k-means. Therefore, the object region is represented as a discriminative appearance model by the vector quantization and the visual codebook. After that, tracking can operate in a Bayesian maximum a posteriori framework. Each candidate regions are represented as a distribution of codebook in descriptor space, which is compared to the template target model. The experiments demonstrate that the superior performance of our method on infrared object tracking. Moreover, the abundant experiments on infrared object in complex background on the sea have demonstrated that our method can track object very well under large appearance change, high speed motion, and partial occlusion.In addition, we propose a novel action recognition method based on the space time points and bag of features. Firstly, local three-dimensional cubical patches within a video are randomly extracted and represented as invariant descriptors by using histogram of oriented gradients and optical flow histogram. Secondly, we make the bag of words model and training and predicting by the support vector machine. The experiments demonstrate that the superior performance of our method on action recognition.