Research on Multitask Model Introducing Self Attention Mechanism and Random Noise Learning in Autonomous Driving

Abstract The multitask models currently used in visual based autonomous driving scenarios face difficulties in model convergence, high training costs, and poor robustness in the process of learning sample features. This paper proposed a new multi task visual model in autonomous driving, which introduced a self-attention mechanism and integrated denoising modules. The model used convolutional structure and self-attention mechanism to extract features from input images, and added denoising elements to the decoder structure of the backbone network. To achieve denoising in various tasks, random noise and simulation experiments were conducted to simulated the distribution of real data. The experimental results show that the new model performs well on multiple autonomous driving tasks, with higher accuracy and inference speed than traditional models. For example, the accuracy of target detection tasks had been improved by 4.3%, the segmentation accuracy of road scenes had been improved by 3.6%, and depth estimation could be performed. At the same time, the training time had been reduced by 60%. This research could provide new ideas and methods for multi task learning in the field of autonomous driving.