Low-latency Block-wise Object Detection Method using SSD for High Resolution Video

In recent years, in the fields such as surveillance cameras and in-vehicle camera systems, efficient deep-learning-based object detection methods, such as Single Shot MultiBox Detector (SSD), that do not require window scanning have received a significant attention. However, these methods require a lot of memory and computation. For this reason, when we applying them to higher definition video, it can be necessary to divide the video into multiple blocks for inference processing due to restrictions on memory capacity of GPUs or FPGAs. To avoid accuracy degeneration due to block division, we can use conventional overlapping block technique. However, it increases the latency of object detection because we need to process more blocks. In this paper, we propose a low-latency block-wise object detection method which assigns a different block pattern into each frame, divides each frame based on the block pattern assignment, and integrates the results of multiple frames. In the experiments, the object detection accuracy and latency were evaluated using three data from the Multiple Object Tracking Benchmark dataset 2017. When the movement of object is small, we reduced the latency of human detection by about 40% while the accuracy degeneration is 0% to 2%.