A Supervised Sensors Based Learning Methodology to Recognize Traffic Signs over Low-Cost Vehicular Sensor Network Environment

The paper presents a supervised learning-based traffic sign recognition system that can be deployed through a low cost vehicular sensor network system. By taking advantage of embedded hardware e.g. the Raspberry Pi 4 with interfacing camera modules, the system conducts real-time classification of traffic signs with a lightweight MobileNet model and sensor fusion methods. Preprocessing and combination of contextual sensor information, in terms of GPS and IMU modules, deal with real-world deployment problems like low light, motion blur and occlusion. The model is trained by means of the Kaggle traffic sign dataset with augmentation with live samples and reaches a score of 97.5 in all clean sign images, and 90-plus in unfavorable conditions. Major contributions are refined edge inference with Tensorrt, a decrease in the energy usage, and enhance strength in changing conditions. Comparative performance of the model with classical models, such as the Random Forest and SVM displays a tangible benefit in both the recognition accuracy and inference latency. The solution can be deployed in intelligent transportation systems in a scalable manner, in particular in cost-sensitive areas. This study shows that it is possible to implement a high-performance and energy efficient sign recognition pipeline with affordable hardware and effective learning techniques and support real-time safety applications in connected vehicular settings.