ZHU FU-ZHEN1, HAN HAO1, JIA HENG-FEI1, ZHU BING21.College of Electronic Engineering, Heilongjiang University, Harbin , 150080
DOI:
Author:
Affiliation:

Heilongjiang University

Clc Number:

Fund Project:

This work is supported in part by the National Natural Science Foundation China (61601174),in part by the Postdoctoral Research Foundation of Heilongjiang Province (LBH-Q17150),in part by the Science and Technology Innovative Research Team in Higher Educational Institutions of Heilongjiang Province (No. 2012TD007)

  • Article
  • |
  • Figures
  • |
  • Metrics
  • |
  • Reference
  • |
  • Related
  • |
  • Cited by
  • |
  • Materials
  • |
  • Comments
    Abstract:

    The current infrared image pedestrian detectors have problems with high rates of false positives and false negatives. To solve these problems, we proposed an improved anchor-free FCOS object detection algorithm. Firstly, we introduced the channel attention module SE-Block in the FCOS backbone network, which was used to learn how to model the relative importance between different feature channels, and to achieve the weight recalibration of the features extracted from the convolution neural network, and improve the weight values that are more important for pedestrian target detection. Secondly, soft non-maximum suppression (soft-NMS) replaced the conventional non-maximum suppression NMS within the algorithm's post-processing phase, which was used to reduce the probability of missed detection for occluded pedestrians. The experimental results show that our improved FCOS algorithm improves the Average Precision(AP) value by 6.71% on the original dataset and 7.97% on the augmented KAIST pedestrian dataset compared with the original FCOS algorithm. Our improvements effectively meet the real-time requirements and there is no significant decrease in speed compared with the original FCOS algorithm, and decreased the false positives and false negatives for infrared image pedestrian detection.

    Reference
    Related
    Cited by
Get Citation
Share
Article Metrics
  • Abstract:
  • PDF:
  • HTML:
  • Cited by:
History
  • Received:July 25,2024
  • Revised:September 20,2024
  • Adopted:October 23,2024
  • Online:
  • Published: