Action Recognition Algorithm Research Based On Sparse Coding Of Oriented Energy Features

Human action recognition is an important topic of computer vision research and applications. The goal of the action recognition is to design a smart system can automatically analyze on-going events in video data. A reliable system capable of recognizing various human actions has many important applications. The applications include surveillance systems, Sports and entertainment system, and a variety of systems that involve interactions between persons and electronic devices such as human-computer interfaces. A vast application prospect inspires more and more scholars engaged in research work in this area. However, human action recognition is a complex project mainly due to some difficult such as the complexity of human action, the large amount of data to be processed, the change of human posture and the conversion of camera perspective. Aiming these difficult, our own action recognition framework will be presented in this paper. The framework is based on Local Spatial-temporal Oriented Energy Features. When the local features are extracted, the samples are purified by sparse coding. Finally, we use Additive SVMs to classify the samples. Specifically:(1) In the stage of interesting feature detection, the classic Dollar detector is easily disturbed by noise in the background. In this paper, Dollar detector is expanded to multi-scale in spatial and temporal space. Exactly, the detecting result on multi-scales will replace the detecting result on a single scale. In this way, the detection will be more stable and accurate.(2) In the features extraction stage, the local energy is decomposed through three-dimensional steerable filter in three directions(X, Y, T). We use this method in the local spatial-temporal region for feature extraction. Then a histogram entropy algorithm is proposed to quantify the features.(3) In the action representation step, the classic “bag of words” method has good robustness on the change of body posture. However, the sparse coding has a good capability to grab the essential information in the samples. In this paper, we combine these two methods. First, we use the local features to learn three over-complete dictionary in the direction X, Y, Z respectively. Then three sparse coefficient histograms were formed to replace the original feature descriptor. Finally, sparse coefficient histograms of all feature descriptors are accumulated to represent the whole action in a video. This method has good inhibition to body posture changes and camera movement.(4) In the action classification stage, the traditional SVM classifier is adopted. Through the experiments of different nonlinear nuclear, we chose the additive kernel as SVM kernel function.(5) In the experimental section, we integrate all modules to find out the optimal grid division of interest neighborhood, the adequate number of basis vectors of over-complete dictionary and the appropriate sparsity of sparse coding. And we take the features in each direction as samples to test the recognition rate. The three type of additive kernel functions are also tested.In this paper, the experimental software platform is Matlab R2010 a. PC configuration is: E7400 processor, 2GRAM. In the experiment, the recognition rate of our system can achieve a higher level.