Foreground Gated Network for Surveillance Object Detection

Object detection in surveillance videos is challenging due to the requirement of real-time detection of small objects in presence of motion blurs and over-exposures at night. False positives caused by above challenges on background areas are more critical in surveillance videos compared to general object detection. Existing state-of-the-art object detection methods often suffer from performance declines, or become time consuming by heavily relying on the temporal correlations between sequential frames. In this paper, we propose a novel real-time method for object detection in surveillance videos to address the above challenges. The proposed architecture employs an efficient BS-RNN block to distill foreground and fine-tunes the unsupervised parameters by minimizing the detection loss. A feature gating block follows to acquire feature-level masks to amplify activations on foreground objects while suppressing irrelevant background regions. We also introduce a new type of feedback connection from object detections to the BS-RNN, which plays an important role in handling static foreground objects such as cars and pedestrians waiting in front of traffic lights. Extensive experiments demonstrate that Foreground Gated Network (FGN) outperforms the compared methods in both accuracy and speed.