Crowd Scene Analysis Encounters High Density and Scale Variation.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Accurate crowd location is crucial for identifying high-risk areas in crowd scene analysis.
• Existing methods struggle with precise localization in highly crowded environments.

Purpose of the Study:

• To propose a novel Compressed Sensing based Output Encoding (CSOE) scheme for improved crowd counting and localization.
• To enhance the robustness of deep models against training errors in crowd analysis.

Main Methods:

• Developed a CSOE scheme that reframes pixel coordinate detection as signal regression in an encoding space.
• Derived a novel sparse reconstruction backpropagation rule to prevent gradient vanishing and enable end-to-end training.
• Validated the method on four mainstream crowd analysis datasets.

Main Results:

• Achieved state-of-the-art performance across all tested datasets.
• Demonstrated superior results in highly crowded scenes compared to traditional pixel-space detection.
• Showcased increased robustness to deep model training errors.

Conclusions:

• Regression in CSOE space is more effective than pixel-space detection for highly crowded scenes.
• The proposed CSOE scheme and backpropagation rule offer a robust and accurate solution for crowd counting and localization.