Adaptive Pedestrian Detection System Based on Deep Learning
No Thumbnail Available
Date
2025-03
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Kenyatta University
Abstract
The existing pedestrian detection algorithms have the potential to improve road safety on a
regional level, however their effectiveness in dynamic rural and urban environments remains
unexploited. With this potential capability, their efficacy remains uncertain due to
infrastructure and operational limitations. Nationally, integration into Kenya’s transport system is still in its infancy, with challenges in policy, infrastructure, and technological readiness limiting real-world deployment. The problem lies in the inability of current systems to provide accurate and timely detection, particularly in complex road topologies such as Type-S roads with sharp curves and frequent occlusions. To address this, this research proposes a YOLO-APD network to enhance detection accuracy and achieve real-time processing. A cost-effective RGB camera in the CARLA simulator was used to generate a custom dataset reflecting diverse traffic scenarios. Enhancements to the YOLOv8 baseline include a novel SimSPPF module for improved feature extraction and speed, a modified detection head with a gather-and-distribute mechanism, and C3Ghost modules for balancing efficiency and accuracy. The model was evaluated through ablation experiments, algorithm comparisons, and robustness tests. Results show YOLO-APD achieved a mean Average Precision (mAP) of 97.8%, with pedestrian detection exceeding 99.5%, outperforming state-of-the-art models. The model demonstrated robust performance with a 94% F1 score, validating its generalization ability in challenging environments. By enhancing detection accuracy and efficiency in Type-S roads, YOLO-APD presents a viable solution for improving autonomous navigation in complex traffic environments.
Description
This Research Project Submitted In Partial Fulfilment of the Requirements for the Award of
The Degree of Master of Science in Computer Science in the School of Pure and Applied
Sciences of Kenyatta University
Supervisor
John Kandiri