A Sparse Spatiotemporal Transformer for Detecting Driver Distracted Behaviors

At present, the development of autonomous driving technology is still immature, and there is still a long way until fully driverless vehicles. Therefore, the state of the driver is still an important factor affecting traffic safety, and it is of great significance to detect the driver’s distracted behavior. In the task of driver distracted behavior detection, some characteristics of driver behavior in the cockpit can be further utilized to improve the detection performance. Compared with general human behaviors, driving behaviors are confined to enclosed space and are far less diverse. With this in mind, we propose a sparse spatiotemporal transformer which extracts local spatiotemporal features by segmenting the video at the low level of the model, and filters out local key spatiotemporal information associated with larger attention values based on the attention map in the middle layer, so as to enhance the high-level global semantic features. Experiments are conducted on a public driver behavior detection dataset (Drive&Act), and the generalization ability of the proposal is evaluated with a dataset collected. Results show that the sparse spatiotemporal transformer devised in this study can obtain robust global semantic features via retaining key local spatiotemporal information while reducing the computational burden, and therefore achieves a high accuracy for the driver distracted behavior detection.