VCrash: A Closed-Loop Traffic Crash Augmentation with Pose Controllable Pedestrian

Collecting and annotating traffic scene datasets is time-consuming and laborious. Virtual data can serve as an important supplement to the real-world traffic scene. Therefore, we propose a closed-loop data augmentation method based on car crash sequences in traffic scenes. Through the CARLA simulation platform, we obtain crash data that is difficult to collect in real-world traffic scenes, and this is then taken as the basis for subsequent pedestrian pose augmentation. To generate rare pedestrian poses, we specifically used boundary value sampling and ideas from genetic algorithms, which broadened the variety of pedestrian stances in the scenes. An efficient detection model was also employed to provide feedback signals for the closed-loop architecture. Based on this feedback, we created scenes for the corner case with various weather, city, and time conditions. The augmentation efficiency can be significantly improved by closed-loop construction. This approach enables us to produce a diverse and comprehensive dataset that can enhance the performance of pedestrian detection models in various traffic scenes.