| Human activity recognition has been one of the most popular topics in computer vision research area. Its applications include human-computer interaction, intelligent monitoring systems and public security systems. A lot of related algorithms had been proposed recent years, but most of them cannot perform great with translation, rotation and illumination changes. For this reason, this thesis studies human action algorithms based on Space-time features including global features and local ones. The major contributions are listed as below:(1) Improvement of the classical global feature based on MEI-MHI algorithm, the original algorithm uses Mahalanobis distance as the classifier which runs fast but cannot achieve a satisfying result, a higher recognition rate is obtained in this thesis by using a SVM classifier.(2) Two local space-time feature points extract method:3D-Harris and Dollar are studied in the thesis, an improved Dollar method was proposed to solve the information redundancy problem in the original Dollar algorithm.(3) Descriptors were formed using gradient of the space-time feature points. And also3D-SIFT algorithm was applied to extract the descriptor which is illumination, translate and rotation invariant. Combined with the above descriptors, bag of words and PLSA algorithm are used in the recognition part, which improved the recognition results.All our work are done with MATLAB R2009b vision on PC. All the algorithms are tested mainly on KTH database, the results prove that our method performs better than others. |
