Research On Vision-Based Human Motion Behavior Recognition

Vision-based human motion behavior recognition based on computer vision and patternrecognition, has great research value for its widely used in the intelligent video surveillance,human-computer interaction, content based image retrieval, virtual reality, motion analysis andother fields. Video-based human motion behavior research is an extremely challenge for itscomplex scene and variety of motion features. Based on a detailed inspection of the relevantliterature, we have a deeply research on human motion detection, motion feature extraction,human motion behavior recognition. The main work is as follows：1. In the aspect of human motion detection, background subtractions based on codebookmethod was used to achieve quick and accurate human motion detection in complex background.To solve the shadow affection caused by the strong light in human motion detection, weproposed a new shadow elimination method base on the OTSU, which can remove the shadoweffectively without threshold selection and the conversion between different color spaces.2. In order to overcome motion feature extraction on the single scale is not comprehensive.We proposed a multi-scaled motion feature extraction method based on human motionmulti-scale classification in the aspect of motion feature extraction. First we choose the Zernikeimage moment features on the intermediate scale as the main feature which has excellent abilityin image express. Then, the trajectory of human motion was improved and used as the seniorscale feature. At last, we combined the intermediate scale and senior scale feature together as themotion feature vector.3. Multi-class support vector machine approach (DAG-SVMs) was used for the humanmotion behavior recognition. We compared the human motion based on single scale (Zernike)feature extraction method and multi-scaled feature extraction method. Experimental resultsshowed that multi-scaled feature extraction method has a higher recognition rate in humanmotion behavior recognition; while DAG-SVMs have a good performance in small samplesituation.4. According to the above research, a vision-based human motion behavior recognition system was designed and the corresponding modules were introduced and tested. The resultsdemonstrate the advantages of the method adopted in the thesis and the effectiveness of thesystem.