Pedestrian Detection from Low Resolution Public Cameras in the Wild
Martin Rajnoha, Lukáš Povoda, Jan Mašek, Radim Bürget, Malay Kishore Dutta · 2018
Since security situation in the world is changing, monitoring of protected areas using surveillance systems has been of increased significance in the recent years. Although the today object detection methods significantly improved accuracy, for real situations, where the video is stream basically of a low resolution and objects are often small and blurry, the methods are still struggling with precise detection. The key parts of any security system are 1) person detection and then also 2) person recognition, which must perform in real-time processing. This paper deals with pedestrian detection in so called wild - i.e. from sources with bad quality, blurry images or small objects for detection. We used Single Shot MultiBox Detector (SSD) which was trained on VOC 2007 dataset and using fine-tuning it achieved percentage increase 11.98% of accuracy for life scenarios. Thus, SSD confirmed its state-of-the-art position and ability to be simply adapted to specific cases of detection while keeping its high performance.