Crowd counting at the edge using weighted knowledge distillation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual crowd counting research has advanced, addressing scale variations and occlusions.
• Existing methods often prioritize accuracy over model size and computational efficiency.
• Resource-limited devices like drones require lightweight, real-time crowd counting models, but these often lack accuracy.

Purpose of the Study:

• To address the accuracy degradation in lightweight crowd counting models.
• To improve the learning capability and generalization of shallow crowd models.
• To enable efficient real-time crowd counting on edge devices.

Main Methods:

• Proposing knowledge distillation to transfer knowledge from deeper models to lightweight ones.
• Training lightweight crowd models to emulate the behavior of larger, more complex models.
• Conducting extensive experiments with three lightweight models across six benchmark datasets.

Main Results:

• Demonstrated significant improvements in the accuracy of lightweight crowd counting models.
• Validated the effectiveness of knowledge distillation for enhancing shallow network performance.
• Ablation studies confirmed the contribution of the proposed knowledge distillation approach.

Conclusions:

• Knowledge distillation is an effective technique for boosting the performance of lightweight crowd counting models.
• The proposed method enables accurate real-time crowd counting on devices with limited computational resources.
• This research contributes to the development of practical crowd analysis systems for real-world applications.