Parallel Attention for Multitask Road Object Detection in Autonomous Driving

Deep learning-based road object detection methods have been widely studied, actively promoting the development of autonomous driving. However, most methods handle tasks individually, which is more time-consuming than parallel processing. Additionally, due to the differences in task characteristics, existing multitask detection methods cannot ensure excellent performance for each task, and there is a problem of unclear boundaries between tasks. To address these challenges, we propose a new unified neural network framework named BHF-MTM, effectively meeting the detection requirements of multiple tasks. Firstly, different shared layers are allocated according to the different characteristics of visual tasks to improve model performance. Furthermore, we design a pixel attention enhancement module (PAEM) to enhance the model’s ability to capture pixel features and spatial position perception. In addition, for tasks that emphasize pixel information, we assist in constructing a segmentation decoder by guided edge refinement module (GERM) to enhance the ability to capture the edge contour clues that were ignored during execution. We evaluate the proposed method on the BDD100K dataset and compare it with existing models. The experimental results indicate that our proposed BHF-MTM has achieved state-of-the-art performance.