Double Attention Convolutional Neural Network for Driver Action Recognition

Driver action recognition is an important technology of Advanced Driver Assistant System (ADAS). Recently, distracted driving recognition with deep learning, especially Convolutional Neural Network (CNN) is very popular, however, the feature extracting in traditional CNN is very sensitive to the true circumstance of driving. It will take a lot of time to train the CNN model with huge amounts of data to gain better robustness. To solve this problem, we introduce the attention mechanism of human visual perception system, which can identify outstanding image regions and magnify their effects while suppressing irrelevant and potentially confusing information of other regions. In this paper, we propose a novel Double Attention CNN (DACNN) to pay attention to salient regions to recognize distracted driving, which combines the local features with the global features to build an attention model, and we use two attention sub-modules to focus on the head, arms and hands of driver, respectively. Our experiment result shows that the recognition accuracy and robustness of our proposed method are improved comparing with CNN without attention sub-module and with one attention sub-module.