A Lightweight Multi-Task Learning Model for Scene Perception in Autonomous Driving

With the support of autonomous driving technology, the development of electric vehicles has accelerated significantly in recent years. Computer vision plays a crucial role in scene perception and decision-making processes in autonomous driving systems. However, efficiently processing vision tasks like drivable area and lane line segmentation in real-time remains a major challenge in resource-constrained environments. To address this aspect, we present DriveLaneNet, a lightweight multi-task learning model optimized for such conditions. The model adopts MobileNetV3 as its backbone and integrates an efficient neck module to merge multi-scale semantic features. Additionally, we introduced an attention module that effectively enhances the model's understanding of global information, improving its performance in complex scenarios. Experiments reveal that DriveLaneNet achieves competitive results in drivable area and lane line segmentation tasks, requiring just 0.3M parameters. On an NVIDIA GeForce RTX 4060 GPU, it delivers over 100 FPS. Although slightly trailing state-of-the-art (SOTA) models in accuracy, DriveLaneNet offers remarkable computational efficiency and real-time capability, making it ideal for embedded applications.