Automatic helmet-less biker detection using deep learning

Abstract

When riders don’t wear a helmet while driving a motorcycle, they aren’t paying attention, leading to crashes and more deaths. The researchers utilized multiple deep-learning models to identify motorcyclists without helmets. We have to identify correctly if a rider is wearing a helmet as we want to reduce risks regarding not wearing a helmet. Our paper proposes a convolutional neural network-based way to decide whether the rider carries a helmet. We utilized pre-trained deep learning models to forecast the outcome because we used a customized dataset. These models include EfficientNetB0, Inception, ResNet50, VGG16, and VGG19, and the results are satisfactory. Later, we put forth our model, combining the CNN, LSTM, and attention models. Our fusion model’s foundation is a Dialated CNN layer. Our dilated CNN layer comprises three maximum pool layers and five convolutional layers. LSTM and an attention model layer follow convolutional layers on a fivelayer CNN. Additionally, we used the same model to predict three classes on a separate dataset, and both models produced satisfactory outcomes. Our goal is to make greater use of the deep learning technique so that it can detect with incredible speed and precision. The test results indicate that, with a classification accuracy of 92.41%, our proposed method outperforms the alternatives we used. We have used YOLOV8 to detect riders wearing non-helmet headwear, such as caps, hijabs, and turbans, and have classified them as non-helmet wearers with satisfactory results.