Aiming at the problem of false and missed detection due to the varied and dense target scale changes, the presence of occlusion, and insufficient light in the target detection task of complex road scenes for self-driving vehicles, an improved model YOLOv8-AUT based on YOLOv8 complex road target detection is proposed. Firstly, the MFR module is designed based on the multi-gradient flow residual structure of the attention mechanism, and the parallel gradient flow branches are added to the module to enrich the gradient flow of the model, so as to enhance the ability to extract the detailed information, and to improve the omission and misdetection of the small targets on the road. Secondly, the spatial pyramid network structure is improved using full-dimensional dynamic convolution to increase the sensory field of the model and improve the model’s ability to detect targets of different scales in complex backgrounds. Finally, the soft-NMS suppression algorithm is introduced to solve the problem of severe target leakage detection in obstacle-target dense regions. The experimental data show that on the BDD100K dataset, the improved algorithm improves the average accuracy mean by 7.7% compared with the original algorithm, mAP@0.5:0.9 by 5.7%, which proves that YOLOv8-AUT can better satisfy the demand for target detection in complex road scenarios of autonomous driving.
Effectively capturing multi-scale object features is crucial for vision sensors used in road object detection tasks. Traditional spatial pyramid pooling methods fuse multi-scale feature information but lack adaptability in dynamically adjusting convolution operations based on their actual needs. This limitation prevents them from fully utilizing spatial hierarchies and contextual information. To address this challenge, we propose a Transformation Dilated Grouped Layer (TDGL) module, a fast multi-scale vision sensor based on deep learning, designed to enhance both efficiency and accuracy in road target feature extraction networks. The TDGL is built upon the Global Layer Normalization Convolution (GLConv) unit, which mitigates internal covariate shift by introducing scaling and offset parameters, modifying dilation strategies, and employing grouped convolution. These improvements enable the network to distinguish features at different scales effectively while optimizing spatial information processing and reducing computational costs. To validate its effectiveness, we integrate the TDGL module into the backbone of several YOLO models, forming the TDGL Net feature extractor. The experimental results obtained on the BDD100K dataset show that the mAP of the TDGL net reaches 40.3% with around 3.1M parameters. The inference speed of the TDGL net after transformation optimization reaches 58 FPS, which meets the requirement for the real-time detection of road obstacle targets by autonomous vehicles.