Abstract:

Driver distraction recognition task can be regarded as a fine-grained image classification task， i.e.， the features contained in a small area of the image determine the category of it. For example， whether a driver is driving normally or chatting with the co-pilot is only determined by the driver’s face orientation. For those categories with slight image differences， the model trained by ordinary image classification method is usually unable to distinguish them with high precision. To solve this problem， a metric space optimized method of distracted driving behavior recognition model training is proposed. By increasing the Euclidean distance between the feature vectors extracted from images of different categories， the model can learn the subtle features to classify these categories， and then improve the model's classification accuracy. The method realizes end-to-end model training without increasing the inference time or introducing in additional supervision information. Experiments on the State Farm dataset show that compared with the ordinary training methods of image classification， the proposed method effectively improves the accuracy of the model.

Key words: driver distraction recognition, metric space optimization, feature vector, image classification