Lightweight Deep Driver Distraction Recognition Network

When the driver shifts his attention away from the driving task and performs behaviors unrelated to driving, it is very easy to slow down the driver's perception ability and cause decision delays. Therefore, timely and accurate identification of driver distraction and issuing of alerts can help improve driving safety. Due to the diverse representations of driver distraction behaviors, it is difficult to collect an dataset which contains all kinds of driver distraction actions to train deep neural networks. This paper proposed an attention-based deep driver distraction recognition neural network DDDNet, which is trained by contrastive learning method to learn to judge complex driving distraction behaviors. Besides, in order to meet the requirement of high processing speed and recognition accuracy, on the basis of the lightweight feature extraction network MobileNetV2, DDDNet proposed to add a three-dimensional channel attention module to strengthen the effective feature. Experimental results show the DDDNet is able to identify a variety of distracting behaviors which it has not met before, with high recognition accuracy of 87.12% and real-time processing speed of 140fps.