Replacing experts by machine in determining the test road grade of autonomous vehicles: Case study in Hefei and policy implications

The road grade classification is a key task for policymakers before autonomous vehicles are tested on the open road. In some pilot cities, classifying the road levels for autonomous vehicle testing requires months of expert consultation. This study seeks to replace the time-consuming process with machine learning methods, using a case study in Hefei to analyze the effectiveness of our approach. We collected data on over 3600 road segments, including road and traffic attributes, and accident records. We constructed several predictive models with various classic machine learning and statistical classification algorithms. We evaluated the strengths and limitations of various algorithms and improved classification performance through data augmentation techniques. Then, we proposed several rules to enhance classification algorithms, effectively mitigating the negative impact of misclassifications in levels 1 and 4. Finally, We tested our algorithms and presented some discussions and the policy implications. Our final integrated method nearly guarantees that segments predicted as level 1 ¹ 1 Road segments available for autonomous vehicle testing at any time. are not of a lower level (we define level 1 as the highest classification), and level 4 ² 2 Roads segments restricted for autonomous vehicle test. segments are not misclassified as other levels. The results indicates that, compared to expert classification, the preliminary classification generated by our method can significantly reduce workload without adversely impacting the traffic system. The methods and results can foster the rapid development of autonomous driving and promote collaboration between enterprises and the government on public welfare projects.