Improving detection speed in video by exploiting frame correlation

Person detection is an important problem extensively studied in computer vision. While most of the approaches are designed for static images, person detection is applied on videos in many real-world scenarios. This is the case for video surveillance and other use cases where the detector remains a significant bottleneck in the system. In this paper we present a principled approach for speeding up the detection algorithm by exploiting the temporal information from the frames. For that purpose, we exploit the inter-frame correlation to accelerate the feature extraction and classification components of the detector. To speed-up the classifier we adapt the cascading threshold by exploiting the frame correlation. The proposed approach provides a speed-up of 5x times while maintaining or increasing the accuracy, or up-to 10x with a negligible drop in accuracy. Furthermore, the proposed approach is generic in nature and can be applied to a large family of methods based on region-proposal and sliding window mechanisms.