Driver Activity Recognition Through Deep Learning

Distracted drivers contribute to a significant proportion of road accidents all over the world. Activities such as texting on cellphones, eating, and reaching for something at the back of the vehicle lead to drivers not paying attention to the road and may result in traffic accidents. This paper proposes the use of residual neural networks (ResNet) with spatio-temporal three-dimensional (3D) kernels to perform distracted driver behaviour recognition. Recently, convolutional neural networks (CNN) with 3D kernels have become an effective tool for action recognition. The 3D kernels extract spatio-temporal features from videos to perform tasks such as human activity recognition. ResNets are a variant of CNNs that utilise skip-connections to realise the training of very deep networks. The large number of parameters in 3D ResNets exposes the possibility of overfitting. Using a large video dataset is thus essential, as it avoids the occurrence of overfitting. This paper examines how different datasets and network depths influence the performance of 3D ResNets. The results are overwhelmingly positive. The findings present a significant positive correlation between the accuracy of a model and the network depth. Furthermore, the quality of the dataset greatly determines the model's ability to generalise effectively.