An artificial intelligence multitier system with lightweight classifier for automated helmetless biker detection

Bike accidents on roads have become a significant concern nowadays. People suffer due to the tragic consequences of road accidents due to the reluctance of bike riders to wear proper helmets, reflecting their lack of awareness, leading to fatalities. This study addresses classification and detection tasks by constructing three custom datasets to identify and categorize bikers riding with or without helmets and assess helmet quality. Photometric data augmentation is applied to balance the dataset images. The primary goal of this study is to develop a computationally efficient convolutional neural network (CNN)-based model named 'BikeNet-12' for accurate classification tasks. Various performance metrics are employed to evaluate the model's overall effectiveness. The BikeNet-12 model achieves a test accuracy of 99.32% in dataset 1. The effectiveness of the proposed approach is validated by experiments on a Safety Helmet Classifier dataset Performance comparison with various transfer learning models demonstrates the applicability of the model in terms of performance metrics. Dataset 2 is utilized to assess helmet quality as safe, unsafe, or inappropriate, and the model achieves the highest accuracy of 98.93%, showcasing its efficacy. Additionally, dataset 3 is employed with the YOLOv8 model to detect non-helmet headwear, such as caps, hijabs, and turbans, among riders, yielding satisfactory results with a mean average precision of 93.7%. Integrating classification and detection tasks positions the model as a potential application to enhance biker safety, promote precaution, and contribute to increased sustainability.