Methodology of Driving Scenario Extraction Based on Vehicle Trajectory

Abstract

In the development of autonomous drive functions, in-house testing and simulation are indispensable, which are very dependent on large amounts of data. Unfortunately, labeling the data and picking up useful scenarios are difficult since there is so much redundant information that is useless. This is why the thesis is born. Specifically, the thesis aims to mine as many interesting driving scenarios as possible with the help of rule-based methods and clustering methods and to find a better way to do scenario extraction from data. Based on the open-source data set NuScenes, the rule based method is first used but with less than satisfactory results due to too little information. The total number of objects cut in, fast approaching, and object cut-out scenarios is around 81, which is relatively small compared to the entire data set, and the classification precision rate is about 50%. Then, clustering models including HDBSCAN, K-means, GMM, Hierarchical clustering, and DBSCAN are tried. On the basis of these models, the work done includes but is not limited to cross-comparing the impact of distance matrices obtained from DTW and LCSS, different partitioning settings, and whether to scale the input data. Certainly, hyperparameter adjustment is also done by the Optuna algorithm. The results from the rule-based method are used to generate metrics and objectives for clustering models. Consequently, the best 𝐹2-score is from the K-means model, which is 0.1688. For comparison, the same method and idea are used on another data set named NGSIM. It turns out that the rule-based results are very good, with the precision rate reaching up to 90% because of the rich information provided like lane ID. Correspondingly, the clustering results improved a lot, reaching 0.4653 for the DBSCAN model, even though it is still far from the ideal result. In summary, the external lane and object information is quite crucial for rule-based method classifying scenarios. Besides, the rule-based method is more efficient than clustering in scenario classification and clustering models may not be applicable to driving scenario extraction based on relative trajectory classification.