Exploiting Temporal Context for Tiny Object Detection

In surveillance applications, the detection of tiny, low-resolution objects remains a challenging task. Most deep learning object detection methods rely on appearance features extracted from still images and struggle to accurately detect tiny objects. In this paper, we address the problem of tiny object detection for real-time surveillance applications, by exploiting the temporal context available in video sequences recorded from static cameras. We present a spatiotemporal deep learning model based on YOLOv5 that exploits temporal context by processing sequences of frames at once. The model drastically improves the identification of tiny moving objects in the aerial surveillance and person detection domains, without degrading the detection of stationary objects. Additionally, a two-stream architecture that uses frame-difference as explicit motion information was proposed, further improving the detection of moving objects down to $4\times 4$ pixels in size. Our approaches outperform previous work on the public WPAFB WAMI dataset, as well as surpassing previous work on an embedded NVIDIA Jetson Nano deployment in both accuracy and inference speed. We conclude that the addition of temporal context to deep learning object detectors is an effective approach to drastically improve the detection of tiny moving objects in static videos.