A BEV Scene Classification Method based on Historical Location Points and Unsupervised Learning

As autonomous vehicles gain consumer favor, ensuring their operational safety has become a crucial aspect of their usage. For those autonomous vehicles, accurate scene perception and recognition of the environment are paramount objectives for ensuring safe decision-making and control strategies. To ensure proper comprehension of driving scenarios, the development of reasonable scene definition methodologies along with highly efficient and rapid scene recognition technologies has become a focal point of interest among researchers. This paper, based on an analysis of the shortcomings of existing scene definition and classification methods, proposes a novel approach to scene data representation. Furthermore, through data validation using the open-source dataset AD4CHE, a new scene recognition method based on the familiarity of vehicles with different scenes is introduced. The method employs an unsupervised learning approach to construct a model for discerning the complexity of scenes. It achieved by utilizing the historical trajectory waypoints of surrounding traffic participants within a fixed time window as inputs. Ultimately, the complexity classification of scenes is accomplished through a three-tier categorization process using complexity indicators derived from the unsupervised model. The scene definition and identification approach proposed is expected to provide valuable data references for the safety assurance systems of autonomous vehicles in the future.