Implementation of a Lightweight Semantic Segmentation Algorithm in Road Obstacle Detection

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Driving Systems

Background:

• Semantic segmentation is crucial for autonomous driving, enabling collision avoidance by identifying obstacles.
• Current semantic segmentation methods face challenges like complex network depth, large datasets, and real-time processing demands.

Purpose of the Study:

• To develop an improved, lightweight, real-time semantic segmentation network for autonomous driving applications.
• To enhance the accuracy and efficiency of obstacle detection and drivable area recognition.

Main Methods:

• An efficient Image Cascading Network (ICNet) architecture was adapted and improved.
• Multi-scale branches and cascaded feature fusion were employed for rich feature extraction.
• A spatial information network was integrated to improve prior knowledge of spatial location and edge details.
• An external loss function was appended during training to boost the learning process.

Main Results:

• The proposed SP-ICNet model demonstrated substantial performance on the Cityscapes dataset.
• The network achieved real-time performance while enhancing the accuracy of road obstacle detection.
• Experimental comparisons confirmed superior performance compared to existing popular semantic segmentation networks.

Conclusions:

• The lightweight SP-ICNet effectively addresses challenges in real-time semantic segmentation for autonomous driving.
• The model provides accurate obstacle detection and road segmentation, crucial for safe autonomous navigation.
• SP-ICNet offers a promising solution for enhancing the perception capabilities of autonomous vehicles.