Citation

Liu L, Ke C, Lin H, Xu H. Comput. Intell. Neurosci. 2022; 2022: e8924027.

Copyright

(Copyright © 2022, Hindawi Publishing)

DOI

10.1155/2022/8924027

PMID

36483290

PMCID

PMC9722493

Abstract

To address the problem that large pedestrian detection networks cannot be directly applied to small device scenarios due to the heavyweight and slow detection speed, this paper proposes a pedestrian detection and recognition model MobileNet-YoLo based on the YoLov4-tiny target detection framework. To address the problem of low accuracy of YoLov4-tiny, MobileNetv3 is used to optimize its backbone feature extraction network, and the MFF model is proposed to fuse the output of the first two layers to solve the information loss problem, and the attention mechanism CBAM is introduced after strengthening the feature extraction network to further improve the detection efficiency; then the 3 × 3 convolution is replaced by the depth separable convolution, which greatly reduces the number of parameters and thus improves the detection rate, then propose Ordinary data augmentation to efficiently augment the dataset and dynamically adjust the target detection anchor frame using the k-means++ clustering algorithm. Finally, the model weights trained by the VOC2007 + 2012 dataset were applied to the pedestrian dataset for retraining by the transfer learning method, which effectively solved the problem of scarce samples and greatly shortened the training time. The experimental results on the VOC2007 + 2012 dataset show that the average means accuracy of the MobileNet-YoLo model compared to YoLov4-tiny, MobileNet-YoLov4, MobileNet-YoLov3, and YoLov5s by 5.00%, 1.30%, 3.23%, and 0.74%, respectively and have reached the level to realize the landed application.

Language: en