CLRNet: A Cross Locality Relation Network for Crowd Counting in Videos

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Crowd counting in videos is crucial for various applications.
• Existing methods struggle with dynamic scenes and accurate density map generation.
• The need for robust feature representation in crowd analysis is significant.

Purpose of the Study:

• To propose a novel Cross Locality Relation Network (CLRNet) for high-quality crowd density map generation.
• To enhance feature representations by modeling pixel dependencies between adjacent video frames.
• To improve the accuracy and robustness of crowd counting in video sequences.

Main Methods:

• Introduced a Cross Locality Relation Module (CLRM) utilizing an adapted local self-attention mechanism.
• Developed an adaptive cosine similarity for measuring positional relationships, outperforming dot product.
• Implemented a Scene Consistency Attention Map (SCAM) to focus on relevant frame correlations.
• Integrated CLRM in a coarse-to-fine manner for enhanced feature representation.

Main Results:

• CLRNet significantly improves crowd density map quality compared to existing methods.
• The proposed adaptive cosine similarity and SCAM contribute to better feature modeling.
• Experimental results on four public video datasets validate the effectiveness of CLRNet.
• The method demonstrates superior performance in crowd counting tasks.

Conclusions:

• CLRNet offers a more effective approach to crowd counting in videos.
• The novel attention mechanisms and similarity measures enhance spatial-temporal feature learning.
• The proposed method advances the state-of-the-art in video crowd analysis.