A DeepSeek-Based Multidimensional Environment-Contextualized Assessment Framework for Driving Behavior Evaluation

Abstract

Existing driving behavior evaluation methods predominantly employ fixed-weight scoring systems that cannot adapt to heterogeneous traffic environments and lack mechanisms for generating personalized, interpretable feedback. This paper proposes an adaptive DeepSeek-based multidimensional environment contextualized assessment (DSMECA) framework for driving behavior evaluation. Firstly, four key dimensions, including safety, efficiency, comfort, and lane discipline, are adopted in the DSMECA framework as evaluation criteria. Accordingly, a continuous adaptive weight system (CAWS) is developed to dynamically compute dimension weights according to real-time traffic states, thereby eliminating the boundary discontinuity and poor scenario adaptability of conventional methods. Then, a hybrid clustering approach combining K-means++ initialization with support vector machine (SVM) boundary optimization is designed to identify driving styles accurately. Furthermore, a decoupled feedback module based on DeepSeek-V3 is introduced to generate interpretable, personalized natural-language evaluation and targeted improvement suggestions through cross-dimensional semantic reasoning. Experimental results from five real-world trajectory subsets demonstrate that DSMECA produces a markedly more discriminative score distribution than the fixed-weight Static Invariant Decontextualized Assessment (SIDA) baseline adopted on the same dataset, achieving an average increase of 22.66% in the highest evaluation tier. Quantitative analysis based on multiple driver samples further verifies the consistency of the LLM-driven feedback module, revealing the application potential in driver training, intelligent traffic monitoring, and traffic safety improvement.