A Novel Algorithm for Efficient Labeling and Its Application to On-Road Risk Detection
Yixin Hu, Qiangfu Zhao, Yoichi Tomioka · 2019
This paper reports on the application of deep learning for on-road risk detection. In the proposed system, a USB-camera is mounted on a mobility scooter to provide video data in real-time, and a convolutional neural network (CNN) is used to detect possible risks. Each frame is classified to 11 categories, including normal, left/right attention, left/right warning, etc. A bottleneck problem in deep learning is the collection of labeled data. During the initial experiment, we collected video data containing more than 130,000 frames to train the CNN. A great number of data will be needed in the process of commercializing the system. To solve this problem, we propose a novel method that enables us to assign labels efficiently. Using this proposed method, we found experimentally that we can obtain labels of all data by labeling manually less than 10% of all the data. In addition, the CNN obtained via transfer learning, based on the well-known AlexNet, performs very well. The average performance of several runs is about 95.83% for testing data. Considering that 30 frames are captured in each second, this accuracy means that three consecutive mistakes are almost impossible, if we use the CNN for real-time risk detection.