Highway abandoned objects have seriously threatened traffic safety, which may lead to vehicle crashes and block roadways. It is significant for transportation agencies to recognize and clear them as soon as possible. However, it is challenging to detect these abandoned objects because they have various sizes/categories and appear at different locations. Although computer-vision-based object detection algorithms have succeeded in traffic monitoring and autonomous driving, most SOTA algorithms cannot satisfy the demand for highway abandoned object detection due to undetermined target categories. Inspired by impressive progress in contrastive learning and large language models, this paper proposes a two-stage open-vocabulary object detection algorithm for highway abandoned object detection. The first stage is the region proposal, with the pre-trained image encoder presenting all region candidates on multiple scales. Then, the second stage adopts CLIP and computes similarities of embeddings from proposed regions. In contrast to conventional object detection algorithms, this algorithm is not constrained by limited training data categories. The proposed algorithm is validated using real-world highway surveillance videos collected in Nanjing, China, containing varying light conditions and multiple video angles. Compared with SOTA object detection models, the proposed one has shown extraordinary performance and generalization abilities for abandoned object detection, with higher precision and running efficiency, that is, in low-light conditions. This approach is helpful for transportation agencies to provide early warning in time and find accident perpetrators.
Highway Abandoned Objects Recognition Based on Open Vocabulary Object Detection Approach
International Conference on Transportation and Development 2024 ; 2024 ; Atlanta, Georgia
13.06.2024
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Handwritten Address Recognition with Open Vocabulary Using Character N-Grams
British Library Conference Proceedings | 2002
|Research on the lightweight detection network of abandoned objects in freeway based on video
British Library Conference Proceedings | 2023
|