An automated learning method of semantic segmentation for train autonomous driving environment understanding

Abstract

One of the major reasons for the explosion of autonomous driving in recent years is the great development of computer vision. As one of the most fundamental and challenging problems in autonomous driving, environment understanding has been widely studied. It directly determines whether the entire in-vehicle system can effectively identify surrounding objects of vehicles and make correct path planning. Semantic segmentation is the most important means of environment understanding among the many image recognition algorithms used in autonomous driving. However, the success of semantic segmentation models is highly dependent on human expertise in data preparation and hyperparameter optimization, and the tedious process of training is repeated over and over for each new scene. Automated machine learning (AutoML) is a research area for this problem that aims to automate the development of end-to-end ML models. In this paper, we propose an automatic learning method for semantic segmentation based on reinforcement learning (RL), which can realize automatic selection of training data and guide automatic training of semantic segmentation. The results show that our scheme converges faster and has higher accuracy than researchers manually training semantic segmentation models, while requiring no human involvement.