A Pedestrian Detection Algorithm Based on Channel Attention Mechanism
Weidong Li, Shuang Han, Yang Liu · 2021
The main contribution of this paper is to introduce the channel attention mechanism into the feature extraction network, and propose the channel attention mechanism module CA, which realizes the efficient fusion of multi-scale features. The deformable convolution is used to replace the traditional convolution operation, and a new detection head is designed, which can predict the position of pedestrians more accurately than the original detection head. CSP is a pedestrian detection algorithm with high accuracy and fast speed, and its structure is very simple. However, there is still great potential for improvement in multi-scale feature fusion and detection head design. This paper proposes a pedestrian detection algorithm based on channel attention mechanism, which is called CA-CSP. On the basis of the original CSP algorithm, the channel attention mechanism module CA is added, and the original detection head is replaced with a detection head based on deformable convolution. The new annotation is used to evaluate the proposed pedestrian detection algorithm CA-CSP on Caltech pedestrian dataset. On the reasonable setting, using a single Nvidia 1660 GPU, CA-CSP has obtained 3.97% of MR−2, and the original algorithm CSP has reached 4.59% of MR−2. Compared with CSP, CA-CSP has lower MR−2. Therefore, CA-CSP has better performance than the original CSP algorithm.