AI-Enhanced Big Data IoT System for Helmet Detection to Boost Biker Safety

This study presents a system aimed at monitoring cyclists to ensure safety compliance, particularly regarding helmet usage. Leveraging big data, computer vision, and deep learning, the research offers a comprehensive approach. Commencing with the collection of cyclist photographs and annotated files containing helmet bounding box coordinates, meticulous organization of big data and file mapping ensures accuracy for IoT systems in urban environments. The YOLO (You Only Look Once) method is employed for object recognition due to its precision and speed, implemented through TensorFlow and PyTorch frameworks. During model training, annotated dataset coordinates serve as ground truth labels, with model parameters optimized to minimize detection errors and enhance accuracy. Post-training, the model is prepared for object detection and inference on unfamiliar images. Utilizing a single forward pass of the YOLO model, inference generates predictions encompassing bounding box coordinates, class probabilities, and confidence scores for identified items, such as helmets. Post-processing techniques are applied to eliminate false positives, refining detection outcomes for IoT systems in urban settings. Evaluation of the bike surveillance system includes precision, recall, and mean average precision (mAP) metrics, indicating the model's accuracy in identifying helmet-wearing cyclists and monitoring safety compliance. This technology automates bike safety law surveillance and monitoring, potentially enhancing safety enforcement by identifying non-compliant riders. Future research avenues may explore safety gear recognition or address environmental factors affecting biking safety. The article also delves into data augmentation techniques, such as rotation, resizing, and flipping of photos, to diversify training data and expedite training, thereby improving model generalization and resilience. Additionally, optimization methods and hyperparameter adjustments impacting training convergence and efficacy for IoT systems in urban environments are discussed. The study's data preprocessing procedures emphasize the importance of consistently sized photos, annotated file conversion for model training, and meticulous data segmentation into train, validation, and test sets to accommodate the scale of big data.