Skeleton Based Human Action Recognition

With the rapid developmet of computer science, research aboutcomputer vision has been in the limelight. Recognition of human actions invideo as an important branch of computer vision has also been a hot topic.Although there are already varieties of algorithms that can achieve greatresults, the performance of the existing algorithms is still not satisfactory forreal data. Therefore, robust human action reconition algorithms that can beagainst variations of illumination, scale, viewpoint, etc., are demanding forpractical applications.This thesis proposes a recognition algorithm of human actions in videoby adding joint angle accelerations as new features. The main tasks are:(1)Based on the previous algorithms of human action recognition, a featureextraction method based on human skeleton is designed.The basic idea is thathuman action consists of a series of human poses and human poses can bedescribed by human skeleton. Therefore, the combination of joint angles andjoint angle accelerations on the human body are used as features to representhuman skeleton.(2) According to the particularity of the feature extractionmethod, Hidden Markov Model (HMM) is chosed as classification schme.The HMM models are trained by features extracted from the CMU motioncapture database.(3) A recognition system of human actions is implementedusing the above-mentioned algorithm.(4) The system is tested on dataset ofCMU and Weizmann. The recognition results are analyzed.The proposed method is invariant to scale, coordinate system, transitionand rotation. It is a relatively stable recognition algorithm. The experimentalresults show excellent performance. It can achieve an average recognitionaccuracy as93.26%on CMU motion capture database, the corresponding variance of accuracy is3.2410-4. It can also get an average reconitionaccuracy as90%on Weizmann classification database. In order to show theimportance of angle accelerations in action recognition, the recognition resultof method without considering angle accelerations and method consideringangle accelerations are compared, and the latter method achieves4.66%accuracy gain on average. A maximum6.41%accuracy gain can be achievedby the proposed method compared with the method without consideringacceleration on the action “run”.