Abstract
Accurate navigation for cyclists requires detailed maps of specific infrastructure, yet standard autonomous driving datasets often overlook cyclist-relevant traffic signs. This paper presents a hybrid computer vision pipeline designed to learn from extremely limited data. By combining the feature extraction capabilities of Convolutional Neural Networks (CNNs) with the small-sample robustness of Random Forests (RFs), and subsequently mapping the RF logic back into a Fully Convolutional Network (FCN), the system achieves high-speed inference and accurate geospatial localization. Trained on an average of only 15 examples per class, the model demonstrates a mean Average Precision of 0.686 and a median localization error of under 7 meters.
Introduction
While autonomous vehicle research has produced massive datasets like KITTI and Cityscapes, these resources primarily focus on motorized traffic. Cyclists operate in a "data desert" where relevant infrastructure—such as bicycle boulevards, paths prohibited for cars, or specific directional signs—is rarely annotated. Furthermore, the perspective of a cyclist, typically using a single handlebar-mounted camera, differs significantly from the multi-sensor arrays used in professional mapping vehicles.
The primary obstacle to creating comprehensive bicycle maps is the labeling bottleneck. Manually annotating thousands of images for every new urban environment is economically unfeasible. This research addresses the challenge of "few-shot" learning: developing a system capable of recognizing and mapping cyclist-specific infrastructure using very few training examples.
Methodology
The proposed pipeline bridges the gap between the robustness required for small datasets and the efficiency needed for large-scale deployment through a multi-stage approach:
- Feature Extraction: A CNN is pre-trained on the large German Traffic Sign Recognition Benchmark (GTSRB) to learn generic visual features.
- Robust Classification: To avoid overfitting on the small target dataset (averaging 15 samples per class), the CNN's features are used to train a Random Forest. RFs are inherently better at generalizing from sparse data than deep neural networks.
- RF-to-NN Mapping: The trained Random Forest is mathematically transformed into a two-hidden-layer Neural Network, allowing the forest's decision logic to be represented as standard weights.
- Fully Convolutional Transformation: The entire pipeline is converted into an FCN. This enables the model to process full images simultaneously, bypassing the computational overhead of sliding-window approaches.
- Geospatial Mapping: Detections are projected onto a global map by combining bounding box coordinates with smartphone GPS and heading data, utilizing pinhole camera geometry and the haversine formula.
Experimental Results
The system was evaluated using a crowdsourced dataset of 40,000 images, with only 297 manually annotated bounding boxes across 10 classes.
Detection and Localization Performance
Despite the limited training data, the patch-wise classifier achieved a test accuracy of 99.3%. The full detection system reached a mean Average Precision (mAP) of 0.686. Performance varied by sign type; for instance, "Bicycle Path" signs were detected with an AP of 0.994, while more complex "Info signs" proved more challenging due to their lack of distinctive geometric features.
The transformation to an FCN architecture yielded massive efficiency gains, reducing processing time from over 10 hours per image (via naive sliding windows) to just 6.08 seconds. Geospatial localization achieved a median error of 6.76 meters, which was further refined to 5.95 meters by filtering for high-quality GPS signals.
Conclusion
This research demonstrates that high-performance object recognition and mapping are possible even with minimal labeled data. By hybridizing Random Forests and Fully Convolutional Networks, the authors provide a scalable solution for enhancing bicycle navigation maps. Future work will focus on multi-view consistency, aggregating detections from multiple frames and riders to further improve localization precision and system reliability.
Summary updated