Driving Action Recognition Based on 3D Convolution

With the number of motor vehicles rising year by year, traffic safety problems are becoming increasingly serious. In this paper, we study the application of convolutional networks to driving action recognition from the perspective of the driver's action in a car. Due to the lack of current driving action datasets, we construct a driving action video dataset DAD-10 containing the entire upper body of the driver. We compare and analyze the effectiveness of 2D convolution and 3D convolution in processing the driving action video data, and based on this, we design a network structure called M(2-1)D that mixes 2D convolution and 3D convolution. The M(2-1)D network uses residual learning and utilizes convolution splitting to reduce the number of network parameters. The network structure achieves a recognition accuracy of 98.5% on the DAD-10 dataset, which has a very high application value.