An improved detection method for multi-scale and dense pedestrians based on Faster R-CNN

The major bottleneck of pedestrian detection lies on the different scales of pedestrians and confusing overlapped pedestrians. To solve these problems, an improved detection method based on Faster R-CNN is proposed in this work. On the one hand, the region proposals are selected according to the usual size-distribution of pedestrians, which can moderately improve the detecting accuracy for different-scale pedestrians. On the other hand, the Soft-NMS algorithm is employed when pruning the redundant detection boxes. It can help to reduce the miss rate caused by overlapped dense pedestrians. With such strategies, the detection performance for multi-scale pedestrians and dense pedestrians can be improved. Finally, experiments to verify the effectiveness of the proposed method are conducted over the challenging Caltech Dataset. It shows that the proposed method outperforms 3.5% better MAP than original Faster R-CNN at the expense of a minor increase in detecting time.