Traffic Sign Detection in Gaode Maps: Machine Learning Techniques and System Architecture

Data is essential for map services; without it, there is no map. While users only see the visible functions of a map, a massive amount of underlying data supports these features.

Map data is collected using vehicles, bicycles, aircraft, and satellite imagery, with recent crowdsourced collection via smart devices. Rapid and accurate data updates are crucial for user experience, and the first step in updating is traffic sign detection.

This article introduces how machine learning is applied to Gaode's map data production, presenting validated technical solutions that enable fast map updates.

1. Introduction

Traffic sign detection refers to automatically identifying various traffic signs (speed limits, no‑turn, crosswalks, electronic eyes, etc.) in street‑view images. The detection results feed into the map data production pipeline.

2. Definition of Traffic Sign Detection

It specifically means detecting different types of traffic signs on ordinary street‑view images using automated methods.

3. Challenges

The main difficulties include:

Huge variety: hundreds of sign types defined by national standards.

Diverse shapes: triangles, circles, squares, diamonds, octagons, as well as ground markings, electronic eyes, signals, barriers, etc.

Wide color range: yellow, red, blue, green, black, white, etc.

Size variance: from several hundred pixels (e.g., square signs) to a few pixels (e.g., electronic eyes).

Natural scene variations such as occlusion, wear, weather, and seasonal effects further complicate detection. Similar‑looking objects like shop signs or advertising boards introduce noise.

Performance requirements are strict:

High recall and precision: missed detections delay data updates; false detections waste resources.

Throughput: billions of images processed daily demand fast inference.

Scalability: the system must adapt to new sign types without degrading existing performance.

4. Gaode’s Traffic Sign Detection Solution

Although academic research often uses end‑to‑end object detectors (Two‑Stage such as Faster‑RCNN, One‑Stage such as YOLO, SSD), Gaode faces three practical issues:

High labeling cost for hundreds of classes and continuous addition of new types.

Inability to iterate on a single class: business needs higher recall for specific signs (e.g., speed limits) but end‑to‑end models require full‑class retraining.

Training difficulty due to class imbalance and large number of categories.

Gaode selected Faster‑RCNN as the base detector because of its superior performance on small objects and its modular Region Proposal Network (RPN), which supports extensibility.

The detection pipeline is split into two stages: a coarse object detection stage and a fine‑grained classification stage.

4.1 Object Detection Stage

Key strategies to improve recall, speed, and scalability:

Group signs by shape (circle, triangle, square, etc.) and assign a dedicated RPN with tailored anchor ratios and scales.

Apply extensive data augmentation and Online Hard Example Mining (OHEM) to balance class distribution.

Incorporate IoU‑Net and Soft‑NMS to boost detection quality.

Share backbone convolutional layers across RPNs to keep inference time low.

For new sign types, add an independent RPN without affecting existing models.

4.2 Fine‑Grained Classification Stage

After detection, each candidate box undergoes a dedicated classification network to eliminate noise and achieve very high precision. Strategies include:

Separate fine‑classification networks for each major shape group, allowing independent and parallel development.

Choose network complexity based on difficulty of the sign type (simple network for circular signs, deeper network for square signs with text).

Collect and label training data per group, improving data preparation efficiency.

Running multiple models increased GPU memory usage, so Gaode optimized the deep‑learning framework by dynamically allocating buffers and pruning unnecessary backward‑pass operations, reducing memory consumption by over 50%.

5. Results and Benefits

The solution is now in production, meeting the required recall and precision thresholds. Daily image throughput exceeds ten million, and the system achieves near‑real‑time (T+0) map data updates.

6. Conclusion

Traffic sign detection has been successfully deployed within Gaode Maps, dramatically improving data production efficiency and achieving near‑instantaneous map updates. Ongoing work continues to apply machine‑learning techniques to further automate map data creation, narrowing the gap between the real world and digital maps.