Explainable Vehicle Activity Recognition: Qualitative and Quantitative Approaches

Deep learning outperforms traditional methods in vehicle activity classification; however, decision interpretability remains crucial for safety-critical applications. This study presents a novel method for classifying vehicle activities involving pairs of vehicles and interpreting the decisions of the classifier. This study consists of four main components. First, we introduce a new set of quantitative features that describes the relative motion between two vehicles as a continuous trajectory of numerical values. Next, we use Bayesian Optimization to develop an optimal Bi-LSTM network, called VqNet, for classifying vehicle activities. Subsequently, we propose a new approach to interpret the decisions made by the Bi-LSTM model to understand the decision-making process. Finally, we conducted extensive experiments using a large dataset comprising three publicly available sets, containing 7,257 trajectories and nine distinct vehicle activities across various traffic scenarios. Our analysis, comparing eight qualitative and quantitative methods, shows that our approach consistently outperforms the others, achieving 11.44% higher accuracy than VNet and addressing its key limitations. Furthermore, our findings indicate that VNet’s decisions align with human reasoning, whereas VqNet relies on specific patterns within activities for classification, without capturing a broader context. This study highlights quantitative models’ performance, simplicity, and differing explainability from qualitative deep learning approaches.