Skeleton-Based Attention Mask for Pedestrian Attribute Recognition Network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian attribute recognition is crucial for intelligent systems.
• Existing methods struggle with variations in human posture and local attribute extraction.
• Skeleton data offers a robust representation for human pose.

Purpose of the Study:

• To develop an advanced pedestrian attribute recognition network.
• To leverage skeleton data with soft attention for improved feature extraction.
• To enhance robustness against posture variations and overfitting.

Main Methods:

• Utilized skeleton data as a soft attention mechanism.
• Designed attention masks for partial and whole-body focus.
• Integrated an augmented layer for data augmentation to mitigate overfitting.
• Evaluated the model on RAP and PETA datasets using ResNet-50, Inception V3, and Inception-ResNet V2 backbones.

Main Results:

• Achieved improved overall classification performance.
• Demonstrated a mean accuracy increase of approximately 2-3% with identical backbone networks.
• Showcased significant improvements in recognizing local attributes.
• Successfully handled variations in human posture.

Conclusions:

• The proposed soft attention model effectively extracts relevant local features for pedestrian attribute recognition.
• The network architecture enhances robustness and reduces overfitting.
• The approach offers a promising direction for improving pedestrian analysis in complex scenarios.