A deeply coupled ConvNet for human activity recognition using dynamic and RGB images

This work is motivated by the tremendous achievement of deep learning models for computer vision tasks, particularly for human activity recognition. It is gaining more attention due to the numerous applications in real life, for example smart surveillance system, human-computer interaction, sports action analysis, elderly healthcare, etc. Recent days, the acquisition and interface of multimodal data are straightforward due to the invention of low-cost depth devices. Several approaches have been developed based on RGB-D (depth) evidence at the cost of additional equipment's setup and high complexity. Contrarily, the methods that utilize RGB frames provide inferior performance due to the absence of depth evidence and these approaches require to less hardware, simple and easy to generalize using only color cameras. In this work, a deeply coupled ConvNet for human activity recognition proposed that utilizes the RGB frames at the top layer with bi-directional long short-term memory (Bi-LSTM). At the bottom layer, the CNN model is trained with a single dynamic motion image. For the RGB frames, the CNN-Bi-LSTM model is trained end-to-end learning to refine the feature of the pre-trained CNN, while dynamic images stream is fine-tuned with the top layers of the pre-trained model to extract temporal information in videos. The features obtained from both the data streams are fused at decision level after the softmax layer with different late fusion techniques and achieved high accuracy with max fusion. The performance accuracy of the model is assessed using four standard single as well as multiple person activities RGB-D (depth) datasets. The highest classification accuracies achieved on human action datasets are compared with similar state of the art and found significantly higher margin such as 2% on SBU Interaction, 4% on MIVIA Action, 1% on MSR Action Pair, and 4% on MSR Daily Activity.