부산대학교 도서관

This research proposes a novel hybrid framework to extract discriminative features for driver distraction recognition. First, the training images are preprocessed to a specific size as input to the InceptionV3, Xception, and MoblieNet models. Second, we fine-tuned the three models by freezing part of the network layer. By visualizing the heat map and feature visualization of every single model, it is concluded that the feature regions associated with different network models are different. Finally, based on the observation of the single models, a deep feature fusion method is designed to fuse the features of the three models. Experimental results on StateFarm database demonstrate that the proposed method attains competitive recognition performance compared with other popular approaches.