RADNet: a highly and robust dynamic network for object detection in complex road scenes

Abstract Aiming at the problems of low detection accuracy and redundant model parameters in the current road scene detection, a RADNet-based road scene detection algorithm is proposed. In this algorithm, an efficient backbone structure CSPResRepBlock is designed to enable the model to obtain more comprehensive target information. Secondly, a multi-feature fusion progressive pyramid is designed to significantly alleviate the elimination of target feature information during the feature fusion process. Meanwhile, a new decoupling head structure called Efficient Dynamic DCNv2 Head is designed, which significantly improves the model’s sensitivity to deformation or edge blurring features. In addition, a structured pruning algorithm is introduced to achieve network acceleration by automatically analyzing complex structural couplings and correctly removing parameters. Finally, experiments on the BDD100K dataset show that the RADNet algorithm improves the average detection accuracy mAP50 by 6.3% and reduces the number of parameters by 68.6% and the computational effort by 30.3% compared to Yolov8s. The inference speed of the model is 114.1 FPS. The model’s generalizability was evaluated using four additional publicly available datasets. These results show that the RADNet algorithm is effective and superior for the task of autopilot target detection.