TransitNet: A lightweight semantic segmentation network for urban traffic scene understanding

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Existing semantic segmentation networks are computationally intensive and lack cross-domain adaptability.
• Deployment on edge devices and in-vehicle systems is challenging due to large model sizes.
• Diverse scenarios like traffic and remote sensing require robust segmentation models.

Purpose of the Study:

• To propose TransitNet, a lightweight semantic segmentation network with enhanced cross-domain adaptation.
• To improve performance on resource-constrained systems and diverse environmental perception tasks.
• To balance efficiency and accuracy for real-time applications.

Main Methods:

• TransitNet is based on PSPNet architecture, incorporating Rectangular Context Calibration Attention (RCCA) and Bidirectional Fusion Attention (BFA) modules.
• A novel StarNet backbone utilizing star-shaped operations and depthwise separable convolutions reduces model parameters.
• An improved forward propagation mechanism supports multi-scale feature fusion.

Main Results:

• TransitNet achieved 86.98% mIoU on VOC2012, outperforming PSPNet by 1.58%.
• On the LoveAD dataset, it reached 61.04% mIoU, surpassing CM-UNet by 8.87%.
• An mIoU of 76.11% on OST300 demonstrates strong generalization and cross-domain adaptation.

Conclusions:

• TransitNet offers a high-performance, efficient semantic segmentation solution for autonomous driving and remote sensing.
• Its lightweight design and cross-domain capabilities make it suitable for edge deployment.
• The model shows significant potential for fine classification tasks in various imagery datasets.