An efficient pedestrian detection network on mobile GPU with millisecond scale

With a certain number of architectural improvements in deep networks, the real-time object detection on mobile devices especially for human beings has gradually become the basis of many downstream visual applications, such as segmentation, tracking, or high-level object-based reasoning. Therefore, there is an urgent need for the object detection model to have higher performance than a standard real-time benchmark. In our study, we proposed a new light-weighted model with a self-designed feature extractor that is based on Resnet18, and using special anchor and loss scheme to eliminate small-sized persons, according to the practical need of subsequent applications to only detect large-sized persons. At the same time, we put forward a newly-designed weighted-NMS method to remove the highly redundant bounding boxes on the same person. The final result outperforms other tiny networks, for example modified ShuffleNet and PeleeNet with nearly the same speed of 5 ms, by achieving 43.3% mAP on person category of COCO-2017 dataset on MI8 GPU with at least 90% in detection accuracy.