A Data Augmentation System for Traffic Violation Video Generation Based on Diffusion Model
Shih-Yu Sun, Ting-Hao Hsu, Chia-Yen Huang, Cheng-Han Hsieh, Chun‐Wei Tsai · Procedia Computer Science · 2024
Intelligent traffic system (ITS) is an advanced application that provides various automatic traffic management services for a city or even a state. The detecting systems in an ITS are usually implemented based on neural network models with the supervised learning algorithm. However, traffic law violation videos in training data are relatively rare which might degrade the accuracy of such a system. Finding high-quality violation data requires considerable effort which makes training a good model time-consuming and expensive. The design of the proposed system is to augment the violation data of the traffic videos with a diffusion model to further provide a good training dataset for the abnormal behavior detection system. The data augmentation of the proposed system can be divided into three parts: (1) the system will first leverage the you only look once (YOLO) model to locate the cars, use an autoencoder to clean the road image and discretize the trail to generate the car's abnormal behavior on the cleared road, (2) the diffusion model will be used to generate images of abnormal behavior based on the position, size, and color of the car, and (3) the images are used to generate the video of the abnormal behavior of the car. The result shows that this system can produce steady-quality traffic law violation videos for object-detection systems. In our experimental results, YOLO is used as the detecting system. The source code of the proposed system is available at: https://github.com/XDOwO/RoadVideoLatentDiffusion.