NAS-PED: Neural Architecture Search for Pedestrian Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian detection in crowded scenes faces challenges like occlusion and dense boundary prediction.
• Convolutional Neural Networks (CNNs) excel at local features, while Vision Transformers (ViTs) capture global dependencies.
• Combining CNNs and ViTs is beneficial, but manual design is resource-intensive.

Purpose of the Study:

• To propose the first Neural Architecture Search (NAS) framework, NAS-PED, for automatically designing hybrid CNN-ViT backbones for crowded pedestrian detection.
• To address the limitations of individual CNNs and ViTs in handling occlusion and local feature details in dense pedestrian scenarios.

Main Methods:

• Developed NAS-PED, a framework that unifies transformers and convolutions for an unconstrained hybrid network search space.
• Introduced an information bottleneck-based NAS objective function to optimize information extraction during the search process.
• Formulated transformers and convolutions with various kernel sizes in a compatible format for diverse hybrid network exploration.

Main Results:

• NAS-PED achieved significant improvements on CrowdHuman, CityPersons, and EuroCity Persons datasets.
• Demonstrated absolute gains of 4.0% MR and 1.9% AP over state-of-the-art methods on CrowdHuman.
• The searched backbones consistently improved performance across subsets and outperformed large language-image pre-trained models on CityPersons and EuroCity Persons.

Conclusions:

• NAS-PED effectively automates the design of hybrid CNN-ViT architectures for crowded pedestrian detection.
• The proposed NAS framework offers a more efficient approach to developing robust pedestrian detection systems.
• The method shows strong generalization capabilities across diverse datasets and challenging scenarios.