Metric-based few-shot learning method for driver distracted behaviors detection

Traditional supervised learning method meets problems such as difficulty in data collection and labeling when applied in special tasks such as driver distracted behaviors detection. Classic Prototype Networks are proposed to solve these few-shot learning problems. However, the fixed prototype representation of the Prototype Network makes its detection accuracy susceptible to the influence of support set samples, making it difficult to apply stably. Therefore, we propose a method to obtain adaptive dynamic prototype representation, and introduce a distance scaling loss function to optimize the training of the network, so that the network can be optimized in the direction of shortening the intra-class distance and enlarging the inter-classes distance. Experiments show that the accuracy of our proposed method in the driver distracted behaviors detection dataset Drive&Act is improved by 3.13% compared to the prototype network, and 9.15% improved compared to traditional supervised learning.