改进的多任务道路特征提取网络及权重优化

Improved multi-task road feature extraction network and weight optimization

为应对自动驾驶在复杂道路环境下的挑战,提出了多项任务合作的需求。在自然语言处理及推荐算法领域,利用多任务学习网络已被证明可以减少多种任务耦合情况下的时间、算力及存储使用。由于多任务学习网络的这种特点,近年来也开始应用于基于视觉的道路特征提取方面。本文提出了一种结合FPN网络的解码器头结构,并将其应用于基于YOLOv4网络的多任务学习道路特征提取网络;通过研究多任务权重设置的影响对多任务网络算法进行优化,并在同类算法中验证了权重设置的有效性。在BDD-100K数据集上进行的试验结果表明,本文结构在保证实时性的同时精度也优于同类方法,本文方法为基于视觉的自动驾驶过程中车辆的自主道路感知及高精地图的生成提供了新思路与新方法。

In order to address the challenges of autonomous driving in complex road environments,the need for collaborative multi-tasking has been proposed.In the fields of natural language processing and recommendation algorithms,the use of multi-task learning networks has been proven to reduce time,computing power,and storage usage in multiple task coupling scenarios.Due to this characteristic of multi-task learning networks,in recent years,it has also been applied to visual-based road feature extraction.This paper proposes a decoder head structure combined with the FPN network and applies it to a YOLOv4-based multi-task learning road feature extraction network.Additionally,the paper optimizes the multi-task network algorithm through investigating the impact of multi-task weight settings.The effectiveness of the weight settings was also verified among similar algorithms.The experimental results obtained on the BDD-100K dataset show that the proposed structure has better accuracy while still ensuring real-time performance compared to similar methods.This paper's method provides new ideas and methodologies for vehicle autonomous road perception and high-precision map generation in visual-based autonomous driving processes.