MSCNet: Efficient and accurate semantic segmentation of LiDAR data using Multi-scale Convolution.

Area of Science:

• Computer Vision
• Robotics
• Sensor Data Processing

Background:

• Semantic information from LiDAR is vital for autonomous systems.
• Processing raw LiDAR point clouds is computationally intensive.
• Existing projection-based methods often have high complexity, hindering real-time use.

Purpose of the Study:

• To propose MSCNet, a computationally efficient multi-scale semantic segmentation method for LiDAR data.
• To improve segmentation accuracy while reducing model complexity and parameters.
• To enable real-time semantic understanding in autonomous driving and robotics.

Main Methods:

• Introduced a single-channel multi-scale feature fusion block to handle input channel differences.
• Employed multi-scale dilated convolution residual blocks for stable local feature extraction across various receptive fields.
• Integrated a pyramid pooling module for rapid global feature acquisition.

Main Results:

• MSCNet demonstrates a favorable balance between model parameters, segmentation accuracy, and processing time.
• Achieved superior performance on the SemanticPOSS and Pandaset datasets.
• Outperformed existing point cloud and projection-based methods under similar parameter constraints.

Conclusions:

• MSCNet offers an effective solution for real-time semantic segmentation of LiDAR data.
• The proposed method enhances efficiency and accuracy for autonomous applications.
• MSCNet provides a competitive alternative to current state-of-the-art approaches.