Improved SSD-Based Multi-scale Pedestrian Detection Algorithm

Abstract

This paper proposes an improved pedestrian detection algorithm based on SSD (Single Shot MultiBox Detector) model. The algorithm mainly solves the problem that the detection effect on small-scale pedestrians is not ideal when the SSD pedestrian detection algorithm performs the pedestrian detection task. The algorithm improves the original SSD model by introducing a shallower feature map. The pedestrian detection is carried out by using the characteristics of different output layers in the model, and the multi-layer detection results are combined to improve the detection effect of small-scale pedestrian. The Squeeze-and-Excitation Module is introduced in the additional feature layer of the SSD model. The improved model can automatically acquire the importance of each channel by self-learning, according to which useful features can be enhanced and features that are not useful for current tasks can be suppressed, thereby further improving the detection ability of the algorithm for small-scale pedestrians. Experiments show that the accuracy of the proposed algorithm in the INRIA dataset reaches 93.17%, and the missed detection rate is as low as 8.72%.