LBT-YOLO: A Lightweight Road Targeting Algorithm Based on Task Aligned Dynamic Detection Heads

Autonomous driving technology plays a key role in addressing traffic safety issues and relieving traffic congestion by virtue of its capabilities of enabling accurate environmental perception and real-time response. Aiming at the problem of limited computing power of mobile driving platform, an improved algorithm based on YOLOv8n: LBT-YOLO was proposed. The algorithm is improved in three aspects: firstly, replacing part of the traditional convolutional layers by linear deformable convolution, and designing a new C2L module by optimizing the C2F module, so as to reduce the number of model parameters and maintain the detection accuracy at the same time. Secondly, a new neck network structure BCFPN (Bidirectional Collocated Feature Pyramid Network) is designed based on the weighted bidirectional feature pyramid network, which enhances the feature fusion and the interaction of contextual information, and improves the detection accuracy of the model. Finally, a new detection head TADDH (Task Aligned Dynamic Detection Head) is proposed. This detection head reduces the number of parameters by sharing the neck network features, and performs task decomposition alignment to achieve high accuracy target detection using dynamic convolution and dynamic feature selection. After a series of improvements, LBT-YOLO outperforms YOLOv8n and other detection algorithms on the Autonomous Driving BDD100K dataset, with an average accuracy improvement of 2.4% while reducing the number of model parameters by 48.2%.