Keyword-Based Vehicle Retrieval

Eun-Ju Park, Hoyoung Kim, Seonghwan Jeong, Byungkon Kang, YoungMin Kwon; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 4220-4227

Natural language-based vehicle retrieval system makes controlling a city-scale traffic system easy to maintain and adaptable to changing requirements. It provides a convenient means in managing traffic flows or detecting accidents related to a specific vehicle. Such a system is different from most query-based video retrieval systems because the language for traffic situations and visible objects in traffic video streams are limited. Existing techniques for language-based general video retrieval problems measure the similarity between language representations and video representations. Our system focuses on several features that can distinguish vehicles from others. Particularly, our proposed vehicle retrieval system defines a set of features that can differentiate a vehicle from others and calculates the similarity between queries and video frames based on the features. The proposed technique places our approach in the third place in the 2021 AI City Challenge.