The application of machine learning models in the analysis of helmet-related images has yielded remarkable results in identifying and classifying helmet-wearing behaviours. Previous research has employed several pretrained models to predict proper or improper helmet use, achieving high accuracy on the Helmet Wearing Image Dataset (2024), a newly introduced dataset designed to enhance classification capabilities. This study aims to improve prediction performance on helmet datasets by leveraging state-of-the-art deep learning models and ensemble techniques. Using ResNet-50, MobileNetV2, and EfficientNet-B0 models, the proposed EnsemHelmet Framework uses a soft voting ensemble to optimise the classification results, achieving an outstanding accuracy of 99.24% on the experimental dataset. The results demonstrate the potential of ensemble learning to achieve high performance. This study not only improves the accuracy of the helmet-wearing recognition system but also highlights the effectiveness of ensemble techniques in optimizing performance on real-world datasets.