Unstructured road recognition is a challenging problem in unmanned driving, involving the complexity of the road itself, such as unfixed type, irregular shape, uneven surface and blurred borders. In order to have a comprehensive understanding of vision-based unstructured road recognition methods and research status, through the analysis and summary of existing literature, this paper analyzes the existing three mainstream methods, which are road features-based, road model-based and machine learning-based methods, and collates the currently commonly used unstructured road open source data sets. The results show that the method based on road characteristics and road model has high computational complexity and low recognition accuracy, and the method based on machine learning can significantly improve the recognition accuracy, but the problems such as large data demand, long training time and poor interpretation are existed as well.

Computer vision  /  Unstructured road  /  Road characteristics  /  Road model  /  Machine learning