Driver-Skeleton: A Dataset for Driver Action Recognition

At present, driver's dangerous driving behavior usually leads to negative outcomes of driving safety and driver action recognition based on skeleton is a current research hotspot. However, there are no large-scale public skeleton datasets for driver action recognition. We present a 3D skeleton information dataset Driver-Skeleton for driver action recognition. This dataset has the advantages of strong pertinence, wide coverage and good scalability. Several experimental subjects are invited to simulate the driver's operation in the cab, and the driver's behavior is divided into 10 classes, which basically covers the driver's common actions in the process of driving. Driver-Skeleton dataset refers to common vehicle models, simulates different vehicle models with different shooting heights, and takes pictures of the experimental objects from different shooting heights. Driver-Skeleton dataset constructed by us used Microsoft Kinect V2 sensor to collect 1423 effective RGB videos from 30 experimental subjects and extract the 3D skeleton information of the driver using these videos. We proposed a two-stream spatial temporal graph convolutional network based on attention mechanism, and experimented on the Driver-Skeleton dataset together with other action recognition methods, and the experimental results confirmed the effectiveness of the dataset. The dataset is freely available at https://github.com/JaxferZ/Driver-Skeleton.git.