LEPD-Net: A Lightweight and Efficient Network for Pedestrian Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Pedestrian detection is vital for autonomous driving and video surveillance.
• Existing methods often prioritize accuracy over efficiency, hindering real-time applications.
• High real-time demands present challenges for practical pedestrian detector deployment.

Purpose of the Study:

• To develop a lightweight and efficient pedestrian detection network (LEPD-Net).
• To address the limitations of current models in terms of computational complexity and inference speed.
• To improve the practical deployability of pedestrian detectors in resource-constrained environments.

Main Methods:

• Designed a PoolFormer-based detection head (PDH) to minimize computation and inference time.
• Developed a triple-branch joint attention module (TJAM) to enhance global context modeling with minimal parameters.
• Integrated PDH and TJAM into a backbone network to create the LEPD-Net architecture.

Main Results:

• LEPD-Net achieved state-of-the-art performance on Caltech and CityPersons pedestrian datasets.
• The proposed model demonstrated a 25% reduction in inference time.
• Accuracy was maintained at a high level despite the efficiency improvements.

Conclusions:

• LEPD-Net offers a significant advancement in lightweight and efficient pedestrian detection.
• The network effectively balances detection accuracy with reduced computational cost.
• This research facilitates the deployment of advanced pedestrian detection in real-time systems.