Abstract
This paper presents a traffic sign detection and recognition method from mobile LiDAR data and digital images for intelligent transportation-related applications. The traffic sign detection and recognition method includes two steps: traffic sign interest regions are first extracted from mobile LiDAR data. Next, traffic signs are identified from digital images simultaneously collected from the multi-sensor mobile LiDAR systems via a convolutional capsule network model. The experimental results demonstrate that the proposed method obtains a promising, reliable, and high performance in both detecting traffic signs in 3-D point clouds and recognizing traffic signs on 2-D images.
Introduction
Maintaining an up-to-date inventory of traffic signs is critical for road safety and the development of autonomous driving systems. Traditional manual inventory methods are labor-intensive and time-consuming. While automated image-based systems have been developed, they are often fragile when faced with environmental challenges such as heavy rain, deep shadows, or varying viewpoints. This research addresses these limitations by fusing Mobile Laser Scanning (MLS) data, which provides accurate 3D geometry regardless of lighting, with digital imagery that offers the necessary color and texture for precise classification.
Methodology
The proposed method utilizes a hybrid pipeline that transitions from 3D spatial detection to 2D semantic recognition.
LiDAR-Based Detection
The first stage involves isolating potential traffic signs within the 3D point cloud. The system performs ground filtering to separate the road surface, followed by object extraction using Euclidean clustering and eigen-based feature descriptors to identify pole-like structures. Because traffic signs are manufactured with highly reflective materials, the system leverages the retro-reflectance (intensity) of the LiDAR returns to pinpoint the exact location of the sign on the detected pole.
Image-Based Recognition
Once a candidate region is identified in 3D space, its coordinates are projected onto the corresponding 2D digital images. To classify the sign, the authors employ a Convolutional Capsule Network (CapsNet). Unlike standard Convolutional Neural Networks (CNNs) that use scalar neurons to represent feature probability, CapsNet uses vectorial "capsules." These vectors encode both the presence of a feature and its properties, such as orientation and size, making the model significantly more robust to changes in viewpoint and minor distortions.
Experimental Results
The framework was evaluated using a dataset collected by a RIEGL VMX-450 system on Xiamen Island, China. The survey covered a route at 50 km/h and captured 1,268 traffic signs across 35 categories.
- Detection Accuracy: The LiDAR-based detection stage successfully identified 1,162 signs, achieving an accuracy of 86.8%.
- Recognition Accuracy: The Capsule Network achieved a high recognition rate of 96.5% on the test set.
- Efficiency: The system is capable of processing 18 traffic sign patches per second, demonstrating its viability for large-scale urban mapping applications.
Conclusion
By using LiDAR as a spatial filter to reduce the search space for image recognition, the proposed system minimizes the false positives common in image-only approaches. While challenges remain regarding occlusions and reflective "imposters" like advertising boards, the combination of geometric precision and capsule-based semantic recognition provides a robust solution for automated traffic sign inventory.
Summary updated