Combined attention mechanism and CenterNet pedestrian detection algorithm

摘要

Traditional pedestrian detection methods can no longer be effective improvement of pedestrian detection, so this paper proposes a pedestrian detection method combined with the attention mechanism, SE-Net attention mechanism and ECA-Net attention mechanism for channel focusing on sex to suppress some not important information channel, which greatly improves the recognition accuracy. Algorithm in this paper, based on CenterNet experiment, by predicting the pedestrian center and return the size information, this kind of anchor-free algorithm greatly reduces the time complexity. By targeting some occlusion and small scale pedestrian attention, our results on the ETH dataset are a good improvement over the original network.