Abstract: In visual perception tasks based on vehicle-mounted cameras, multi-task collaboration has become an important technical paradigm for intelligent driving. However, efficiently extracting multi-task features and achieving feature complementarity in a single inference pass remains challenging. To address this problem, MtTEPNet, a multi-task traffic environment perception algorithm, was proposed for object detection, road drivable area segmentation, and lane line segmentation.First, a dynamic multi-scale feature fusion neural module was designed to integrate hierarchical semantic information and enhance the model’s capability to capture information from complex traffic scenes.Then, a cross-task interactive attention mechanism was introduced to model global correlations between object detection and image segmentation tasks, achieving complementary feature enhancement and reducing mutual interference.Finally, a large-Kernel separable attention module was constructed to extract pixel-level long-range contextual relationships from image features, thereby enhancing feature representation for slender structures and boundary objects.Experiments on the BDD100K dataset show that the proposed algorithm achieves a mean Intersection-over-Union of 92.8% for drivable area extraction, an accuracy of 87.4% for lane line segmentation, and a mean average precision of 79.9% for object detection, with only 8.3 M parameters, 13.8 GFLOPs, and an inference speed of 38 frames per second. The proposed method outperforms mainstream multi-task traffic environment perception algorithms in overall performance.