Action Recognition Based On Shape Context And SURF Interest Points

Human action recognition has been widely used in many fields, such as banks,railway stations, airports, shopping malls, roads, office buildings and residentialareas. Nowadays the computer-based human action recognition technology has beenapplied to motion capture, video surveillance, video classification, sports andentertainment video processing, and intelligent house&home, human computerinteraction, environmental control and monitoring and so on.The main human action recognition methods are generally based on thefollowing three aspects: the human-body-model-based tracking, the histogram ofoptical flow, and local spatial and temporal characteristics. Each aspect has its owninadequacy. The methods based on the human-body-model-based tracking need toextract the accurate human body template. Therefore, those methods have arelatively poor robustness. The methods based on the histogram of optical flow usethe optical flow information for action recognition. But the optical flow is sensitiveto the background noises and different light intensities. The methods based on localspatial and temporal characteristics use filtering to get the human action’sspatial-temporal key points with exceeding value of filter response. For thelow-resolution input image or video, the spatial-temporal key point method also hasa good effect. Moreover, it also has a good robustness to the scale, shooting angleand light. On the basis of the previous studies, our paper will propose two differentmethods: Shape-Context-based and SURF-Interest-Points-based methods.The methods based on Shape Context need to extract the contour of the humanbody. Our method defines a shape distance based on shape context. According to theminimum distance criterion, we cluster certain key contours by using theshape-distance-based K-medoids algorithm for per action. In the recognition process,given an action video, we compute the distance between each frame and all actions’key contours. By using the majority voting rule, the action sequence is classified asthe action whose key frames get the majority matching votes. The experimentalresults have shown that an appropriate number of key frames can well represent anaction. The method based on Shape Context achieves a high classification accuracyon the KTH, UCF and Weizmann action datasets.In order to improve the algorithm’s robustness for occlusion, overlapping, scaleand illumination changes, our method apply SURF interest point to actionrecognition. The SURF interest point has rotation, scale, translation invariantproperty. The methods based on SURF interest points select the features of humanaction from the video. Firstly, we detect the spatial-temporal key points by usingSURF detector. Secondly, we construct the point set of the trajectory by accumulating the interest points both salient in time and space scales, and thenextract the corresponding features with multi-scales for action recognition.Experimental results indicated that the algorithm combining SURF interest pointsand the trajectory information has obvious advantages. The point set of thetrajectory can significantly describe the formation process of human motion. As theSURF interest points do not depend on accuracy extraction and positioning, thealgorithm is robust to background noise and light intensity.