Convolutional Neural Network-based Jaywalking Data Generation and Classification

In this paper, we propose a novel system to generate jaywalking images. To synthesize a pedestrian on the road and label the binary case such as jaywalk or normal-walk, the pre-trained Convolutional Neural Network (CNN) is used to segment the drivable area from the large-scale dataset. The proposed system automatically generates a jaywalker based on existing pedestrian objects in the image. The proposed system performs three main steps. First, we train the existing network with both black box image dataset and object dataset to segment road areas and pedestrians. Second, the generator synthesizes jaywalkers randomly within the road segmentation masks. Third, a CNN classifier is trained using the generated synthetic dataset and performs the inference from natural jaywalking images. The experiment results show that the jaywalking classifier trained with both generated synthetic dataset and the untouched natural dataset has a high accuracy of 0.96, which is 0.08 higher than the accuracy using only the untouched natural dataset on the same model.