A multi-branch separable convolution neural network for pedestrian attribute recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video surveillance is crucial for safety, but extracting visual details like clothing or accessories is challenging.
• Existing convolutional neural network (CNN) solutions for pedestrian attribute recognition often require large datasets.

Purpose of the Study:

• To develop an efficient pedestrian attribute recognition method using a novel multi-branch CNN architecture.
• To leverage depthwise separable convolution (DSC) layers and multiple color spaces for improved performance.

Main Methods:

• A multi-branch CNN, termed 3bCNN, was designed incorporating depthwise separable convolution (DSC) layers.
• The network utilizes different color spaces across its branches to capture diverse visual features.
• Experiments were conducted on two benchmark datasets to evaluate the proposed model.

Main Results:

• The 3bCNN model demonstrated improved performance in pedestrian attribute recognition compared to existing state-of-the-art methods.
• The architecture proved particularly efficient when applied to smaller datasets.
• The use of DSC layers and multiple color spaces contributed to the enhanced results.

Conclusions:

• The proposed 3bCNN architecture offers an efficient and effective solution for pedestrian attribute recognition.
• This method advances the capabilities of video surveillance systems by accurately extracting visual attributes.
• The findings suggest a promising direction for CNN applications in computer vision, especially with limited data.