A Pixel-level Image Synthesis Method Based on Vehicle Position Semantics and Motion Attitude

Computer vision technology is considered a key component in enabling autonomous vehicles to understand their environment. However, training visual perception models typically requires large-scale, high-quality image and video datasets. To alleviate this problem, a semantic image synthesis model based on vehicle pose and position is proposed, enabling the synthesis of perception images with semantic-level labels directly within existing datasets. Specifically, we first utilize the YOLOv8-seg model on the BDD100K dataset to extract foreground images containing vehicle semantic information and obtain high-quality background images from the CUlane dataset. Then, we introduce an innovative multi-semantic standard image synthesis model, which rationally integrates foreground and background images by considering semantic information such as vehicle pose, vehicle behavior, lane markings, and drivable areas. The proposed model effectively enhances the semantic diversity of the original dataset, expanding its size by a factor ranging from 2 to 60. Finally, we apply image harmonization techniques to refine pixel-level consistency, ensuring the realism of the synthesized images. The proposed method is evaluated to validate its effectiveness and superiority.