Vehicle Kinematics-Based Image Augmentation against Motion Blur for Object Detectors 2023-01-0050

High-speed vehicles in low illumination environments severely blur the images used in object detectors, which poses a potential threat to object detector-based advanced driver assistance systems (ADAS) and autonomous driving systems. Augmenting the training images for object detectors is an efficient way to mitigate the threat from motion blur. However, little attention has been paid to the motion of the vehicle and the position of objects in the traffic scene, which limits the consistence between the resulting augmented images and traffic scenes. In this paper, we present a vehicle kinematics-based image augmentation algorithm by modeling and analyzing the traffic scenes to generate more realistic augmented images and achieve higher robustness improvement on object detectors against motion blur. Firstly, we propose a traffic scene model considering vehicle motion and the relationship between the vehicle and the object in the traffic scene. Simulations based on typical ADAS test scenes show that the high vehicle speed and near object position is the key factor in generating motion blur. Second, we propose the vehicle-motion-based image augmentation algorithm. The proposed method applies the motion blur on the clear object based on the vehicle's speed and the relative position of the object. Subjective evaluation and multiple objective evaluation indexes including structural similarity index measure (SSIM), perceptual hash, normalized mutual information, and cosine similarity demonstrates that the proposed image augmentation can produce images more consistent with the traffic scenes. Thirdly, we apply the proposed method to the training of object detectors. Experiments on the KITTI dataset as well as real-world driving tests show that the proposed image augmentation achieves a higher robustness improvement than existing image augmentation algorithms on multiple object detectors including CenterNet, YOLOv3, and Faster R-CNN.