A human head detection method based on center point estimation for crowded scene

The head detector can effectively handle the Angle change of the head caused by motion, and the features required to capture are smaller. Therefore, human head detection is widely used in the practical application scenes of people positioning and counting. The existing head detector uses a large number of anchors which makes the detection efficiency low. And additional post-processing is required, which may result in the loss of real objects in a crowded scene. In this paper, we return to the bounding-box of the head by estimating the center point of the human head without using anchors and post-processing, which improves the detection efficiency. At the same time, we designed a random combination of data augmentation methods and improved the backbone network to improve the accuracy and robustness of the head detector in crowded scenes. Our method achieves excellent speed and precision performance on the SCUT-HEAD dataset, with 0.91 AP and 0.70 EER at 61.5 FPS.