Phase Congruency Based Human Action Recognition And Classification

In recent years, as video databases expand rapidly in size and video surveillance systems continue to proliferate, there is a fast growing demand for automatic video analysis and management technology, because human labor can no longer handle the complexity of such tasks. Among video analysis technologies, the most promising segment is human action recognition and classification. The current mainstream approaches are based on spatiotemporal interest point. However, existing methods of spatiotemporal interest point detection have two important drawbacks. First, the detection response varies significantly under different contrast or illuminations. Second, the application of Gaussian smoothing degrades the accuracy of the feature locations. To address these two problems, we propose a novel phase congruency based spatiotemporal interest point detection algorithm and its corresponding description algorithm.In order to improve feature detector's invariance to contrast and location accuracy, we propose a novel phase congruency based spatiotemporal interest point detection algorithm, extending the concept of phase congruency from image analysis to video domain. The phase congruency measure responds strongly to all kinds of spatiotemporal edges. Compared with existing methods, our approach has two key advantages. First, phase congruency is a dimensionless measure so it is independent of variances in illumination and contrast. Second, phase congruency based method does not apply any forms of Gaussian smoothing, so the location of spatiotemporal interest point is more precise.To enhance the discriminative power of feature descriptor, we propose a novel phase congruency based method to go with the detection algorithm. First we suggest a method to calculate the spatial scale and temporal scale of the interest points. Thus we can extract a cuboid around any given interest point. Then we propose a novel phase congruency based feature descriptor.Compared with traditional grayscale based descriptor, our method can better capture the complex structure of structure and object motion. So this novel descriptor can improve the accuracy of action recognition and classification.For the design of action recognition and classification system, we first argue our choice'bag of words'model as our system framework. Then we discussed the choice of classification algorithm. Finally we discussed the selection of learning data and the configuration of codebook size. Finally, we tested our key conclusions on KTH human motion dataset. Experiments show that phase congruency based interest point detector exhibits strong detection performances and invariance to contrast and illuminations. We also demonstrate that the behavior recognition and classification system based on phase congruency detector and descriptor achieves very high accuracy.