2021 年 76 巻 5 号 p. I_899-I_907
In the research field of video processing, various methods for detecting damaged parts of roads and objects have been developed. However, it is difficult to define all of the obstacles on roads due to the wide variety of obstacles. Therefore, a novel multimodal method for detecting cyclists' avoidance behavior caused by the presence of obstacles using dashcam videos is proposed in this paper. The proposed detection method focuses on cyclist behavior and bicycle motions. The proposed method consists of three stages. In the first stage, features of cyclist behavior and bicycle motions are extracted. In the second stage, the proposed method performs tentative recognition. Specifically, the proposed method recognizes cyclists' avoidance behavior using features of cyclist behavior. Recognition using features of bicycle motions is also performed. Thus, the proposed method obtains two recognized labels. Finally, by integrating the two recognized labels using the corresponding attribution probabilities, the proposed method obtains final detection results. Experimental results show the effectiveness of the proposed method.