Driver’s mobile phone usage detection using guided learning based on attention features and prior knowledge

There is extensive evidence that mobile phone use while driving contributes significantly to driver distraction, increasing the risk of traffic accidents. This work proposes a method to detect a driver’s mobile phone usage (DMPU). We use images of drivers and a deep neural network (DNN) to detect DMPU for efficient and straightforward implementation. However, several problems limit the accuracy and generalization ability of driver behavior detection models and their implementation in vehicles, such as not considering the importance of different areas for feature extraction, numerous driver images for training, and people’s distrust in the prediction results of DNNs. To alleviate these problems, we design a DMPU detection model using guided learning based on attention features and prior knowledge. We label the driver’s images with attention features and extract them with a DNN. An anthropomorphic attention mechanism with 3-D weights based on neuroscience and spatial suppression theories is incorporated as an energy function to calculate the importance of each neuron in the feature map. A guided learning method based on the attention features extracts the most important features of DMPU. The prior knowledge obtained after training the proposed model on an existing classification dataset is used as part of the weight to improve the training results and significantly improve the generalization ability of the driver behavior detection model given insufficient training data. Finally, we propose a method to visualize features related to driving behavior. Experimental results show that our model can accurately detect DMPU.