Driver Action Recognition Based on Attention Mechanism

Abstract

According to the world health organization, millions of people are killed by traffic accidents worldwide every year, and more than 80 percent of accidents are caused by unsafe driving. This paper studies driver behavior recognition, aiming to standardize driver's driving behavior and reduce the probability of traffic accidents. However, the inter-class variance of drivers' different actions is small, making it difficult to identify. To improve fine-grained identification, an attention module is designed to be inserted into convolutional neural network, which consists of two parallel parts: channel level attention and space level attention. The introduction of attention mechanism can focus the weight of the network on the meaningful pixels and channels, promote the expression of effective features, and suppress the interference of noise. The experiments show that the recognition accuracy is improved after applying attention mechanism. The visualization results show that the introduction of attention mechanism can make the network focus on the prominent areas of the feature map.