Abstract
In the area of computer vision (CV), action recognition is a hot topic of research nowadays due to famous applications, which include human-machine interaction, robotics, visual surveillance, video analysis, etc. Many techniques are presented in the literature by researchers of CV, but still they faced a lot of challenges such as complexity in the background, variation in the camera view point and movement of humans. A new method is proposed in this work for action recognition. The proposed method is based on the shape and deep learning features fusion. Two-steps-based method is executed- human extraction to action recognition. In the first step, first, humans are extracted by simple learning process. In this process, HOG features are extracted from few selected datasets such as INRIA, CAVIAR, Weizmann and KTH. Then, we need to select the robust features using entropy-controlled LSVM maximization and performed detection. Second, geometric features are extracted from detected regions and parallel deep learning features are extracted from original video frame. However, the extracted deep learning features are high in dimension and some are not relevant, so it is essential to remove irrelevant features before fusion. For this purpose, a new feature reduction technique is presented named as entropy-controlled geometric mean . Through this technique, we can select the robust deep learning features and remove the irrelevant of them. Finally, both types of features (selected deep learning and original geometric) are fused by proposed parallel conditional entropy approach. The obtained feature vector is classified by a cubic multi-class SVM. Six datasets (i.e., IXMAS, KTH, Weizmann, UCF Sports, UT Interaction and WVU) are used for the experimental process and achieved an average accuracy of above 98.00%. The detailed statistical analysis and comparison with existing techniques show the the effectiveness of proposed method .