Robust Pedestrian Detection for Autonomous Vehicles Using a Fortified Network Model
Part Of
Conference Proceedings - 2025 9th International Conference on Man-Machine Systems, ICoMMS 2025
Start Page
6
End Page
11
ISBN (of the container)
9798331595401
ISBN
9798331595401
Date Issued
2025
Author(s)
Tsai, Yao-Yi
Abstract
Detecting pedestrians in high-density and crowded scenes for autonomous vehicles presents numerous challenges including occlusions, lighting changes, and dynamic human postures. This paper proposes a fortified network framework for pedestrian detection with the main component containing a ResNet-50-based architecture. A self-attention mechanism called the discrete dimension attention module (DiDAM) is incorporated into the ResNet-50 architecture. The DiDAM learns complex features through the channel, height, and width directions to enrich pedestrian representations in dense population environments. Additionally, the residual block in the ResNet-50 model is enhanced by introducing the grouped residual block (GRB), which concatenates group convolutions instead of simple addition to reinforce the feature extraction capability. To assess the effectiveness of this system, the public CrowdHuman dataset was utilized. Experimental results showed that the proposed network model outperformed many competing methods and achieved excellent performance evaluation scores with the average precision (AP) of 90.45%, log-average miss rate (MR-2) of 41.28%, and Jaccard index (JI) of 83.05%.
Event(s)
9th International Conference on Man-Machine Systems, ICoMMS 2025, Malacca, 18 August 2025 - 19 August 2025
Subjects
AI
autonomous system
DiDAM
GRB
grouped convolution concatenation
pedestrian detection
SDGs
Publisher
Institute of Electrical and Electronics Engineers Inc.
Type
conference paper