Selective Scale-Aware Network for Traffic Density Estimation and Congestion Detection in ITS.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Engineering

Background:

• Traditional traffic congestion detection methods struggle with vehicle scale variations in surveillance video.
• Existing approaches often rely on vehicle detection or holistic mapping, limiting effectiveness.
• Solely using spatial features of traffic density overlooks the dynamic nature of traffic flow.

Purpose of the Study:

• To develop an advanced method for accurate traffic congestion detection in surveillance footage.
• To address limitations of traditional methods in handling scale variations and dynamic traffic characteristics.
• To propose a multi-task framework for simultaneous estimation of traffic density and dynamic congestion.

Main Methods:

• Proposed a Selective Scale-Aware Network (SSANet) for generating traffic density maps.
• Integrated static congestion level estimation from density maps using a linear layer.
• Incorporated traffic flow velocity with static congestion for dynamic congestion assessment.

Main Results:

• Achieved state-of-the-art results on both congestion detection and density estimation tasks.
• Demonstrated high accuracy in classifying traffic flow, outperforming existing methods.
• The SSANet model achieved 99.21% accuracy on the UCSD traffic flow classification dataset.

Conclusions:

• The proposed multi-task framework effectively detects traffic congestion and estimates density from surveillance video.
• The Selective Scale-Aware Network (SSANet) enhances accuracy by addressing scale variations.
• This approach offers a more comprehensive understanding of both static and dynamic traffic congestion.