The Use of Vision Foundation Models Combined with Temporal Context for Vehicle Detection

This paper addresses the problem of vehicle detection in intelligent transportation systems based on video data from surveillance cameras. A method is proposed that combines a DETR-like architecture based on DINO, features from a vision foundation model (CLIP), and temporal context extracted from frame sequences. Unlike existing approaches, the model simultaneously utilizes vision foundation model features and temporal context within a single architecture. CLIP features are integrated at both encoder and decoder levels to incorporate semantic information, while temporal context is implemented through aggregation of queries between current and reference frames followed by processing in a dedicated temporal block. Experiments are conducted on the UA-DETRAC dataset with refined annotations. The results show that the proposed method provides higher confidence in predictions and improved robustness to noise, although the improvement in accuracy metrics remains modest. The obtained results demonstrate the potential of the proposed approach for traffic monitoring systems and suggest directions for further research.