D2MM-CNN: DIFFERENCE DEPTH MOTION MAP AND CONVOLUTIONAL NEURAL NETWORKS FOR HUMAN ACTION RECOGNITION

Abstract

Human Action Recognition has become the most significant research area for several applications like robotics, healthcare, gaming, smart houses, etc. However, in computer vision, action recognition from videos is one of the most challenging issues, due to some extraneous aspects like Occlusions, backgrounds, noises and so on. One solution to overcome the above-mentioned problems is acquiring only motion and shape cues form depth action video sequences. With this objective, in this paper, a new action representation approach is proposed based on Depth Motion Map (DMM), called as Difference Depth Motion Map (D²MM). Next, a well-designed CNN is trained especially to extract the features from two actions with a similar structure. The CNN model introduced in this paper involves five convolutional layers, three pooling layers, and one fully connected layer. The experimental results of the proposed method are compared with conventional methods on the publicly available dataset, MSR Action 3D. The comparative analysis proves that the proposed approach outperforms the state-of-art techniques.