Video-Based Bicycle Detection in Underground Scenarios

Abstract

Automatic surveillance systems are an important emerging application of object detection algorithms in video. The nature of such systems implies several requirements on the used algorithms. Also, searching for less usual objects (in contrast to frontal human faces, car masks, etc.) is required, such as detection of bicycles. It appears that detection of such objects cannot be solved by just applying a standard statistical or other general detector, but by constructing a specialized detector composed of several standard image processing and object-detection techniques combined together ad hoc. A detector of bicycles in video data from standard low-resolution CCTV surveillance system is presented in this contribution. Bicycle detection approach covered by this paper aims to cope with highly-noisy low-resolution data, to use simple image-processing methods and to work in real time. Although the method itself does not constitute a generally usable object detector, it covers several interesting aspects which can be re-used in tasks similar to the given one. Low-level features extracted from the video used for wheel-candidate classification are described in detail. The system is applied and evaluated on real data and the results are discussed.