Skeleton-based driver action recognition using ResGGCNN
Roksana Yahyaabadi, Soodeh Nikan · 2023
Driver action recognition serves a crucial role in assessing a driver’s distraction and readiness to takeover control in Level 3 of autonomy. In this study, we proposed an effective, driver-independent graph convolutional neural network (GCNN) using the skeletal representation by comprising determinative joints and inter-joint distances in the driver’s body structure. The proposed technique eliminates superfluous information. To distinguish various action classes, our proposed GCNN leverages residual gated connections (ResGGCNN), which enables the model to learn features efficiently at multiple levels. The proposed ResGGCNN is superbly lightweight with only almost 110k learnable parameters, makes it adequate for low-latency and embedded deployment. We evaluated our proposed model on the 3MDAD dataset with 16 common driving and non-driving related activities and achieved significant improvement in the recognition accuracy of driver’s action at the rate of 59.71% on RGB front-view modality.