Practice-Oriented Real-time Person Occurrence Search System

Face recognition is a potential technology to realize Person Occurrence Search (POS) application which retrieves all occurrences of a target person over multiple cameras. From the industry perspective, such a POS application requires a practice-oriented system that can respond to search requests in seconds, return search results nearly without false positives, and handle the variations of face angles and illumination in camera views. In this paper, we demonstrate a real-time person occurrence search system that adopts person re-identification for person occurrence tracking to achieve extremely low false positives. Our proposed system performs face detection and face clustering in an online manner to drastically reduce the response time of search requests from users. To retrieve person occurrence count and duration quickly, we design a process so-called Logical Occurrences that utilizes the maximum interval of detected time of faces to efficiently compute the occurrence count. Such a process can reduce the online computational complexity from O(M2) to O(M) by pre-computing elapsed time during the online face clustering. The proposed system is evaluated on a real data set which contains about 1 million of detected faces for search. In the experiments, our system responds to search requests within 2 seconds on average, and achieves 99.9% precision of search results over more than 200 actual search requests.