Human Motion Analysis Based On Video

Human motion analysis based on videos has become one of the most popular researchtopics in artificial intelligence and pattern recognition recently. It has a wide range ofapplications in human-computer interaction, virtual reality, intelligent video surveillance,smart home, motion analysis and so on. Due to the characteristics of human motion such asnon-rigid bodies, human body self-and inter-occlusions, complex backgrounds, the variety ofclothing, and illumination changes, the human behavior recognition becomes more difficultand video-based human detection is still facing enormous challenges.This thesis presents the researches on the related problems of human action recognition.They mainly include4aspects: image preprocessing, human motion object extraction, featureextraction and the algorithm of human action recognition. The purpose of the thesis is torecognize such behavior like “walking”,“jogging”,“running”,“boxing”,“hand waving”, and“hand clapping”.The main contributions of this thesis are as follows:1. In the aspect of image preprocessing, several methods of image preprocessing, includingimage enhancement, image denoising and image morphology processing are introduced.In the end, experiments are carried out on the extracted human binary image morphologyprocessing, corrosion, expansion and corrosion with expansion algorithm of binary imageprocessing. Results show that corrosion combined with expansion operation and effective operation not only eliminates background noise isolation, but also fills the hole in thetarget to get ideal detection segmentation result.2. In the aspect of human motion detection, moving object extraction is the research base ofvision-based human action recognition. First of all, preprocessing steps for the videossuch as frame extraction and gray scale conversion are performed. Secondly, based onrelated works, we propose a moving object detection algorithm based on the backgroundsubtraction method and the optical flow method for human binary image extraction anduse mathematical morphology to process the holes and noise points existing in binaryimages. Experiments show that the proposed algorithm, compared to Gaussian mixturemodel, is able to accurately detect human movement in real time. It can also extract thecontour of the human body movement target more completely and has good robustness.3. In the aspect of human motion feature extraction, the mainstream of feature extractionalgorithms for human detection is studied. Methods based on HOG feature extractionalgorithm for grayscale images in videos, image normalization, gradient calculation,gradient direction histograms, normalization, descriptor generation, and featuredescription block are implemented.4. In the aspect of human motion behavior recognition, methods based on the directiongradient histogram (HOG) and the support vector machine (SVM) classification algorithmof human behavior are implemented. At the same time, the algorithm of this paper,compared to the experimental results of the local features (LF) and the support vectormachine (SVM), improves the overall human behavior recognition accuracy and hassolved the low accuracy of running and jumping recognition in related works. It is proventhat the method used in this thesis has good performance and effectiveness.