Traffic police gesture detection by autonomous vehicle driving to act using hybrid architecture combining baseline CNN, MobileNet and DenseNet

Autonomous vehicle (AV) technology development has driven up the need for systems which recognize human signals including traffic police gestures with reliability. A deep learning hybrid framework based on baseline CNN, MobileNetV2 and DenseNet121 demonstrates promising capabilities for accurate traffic police gesture recognition systems. The specialized dataset contains ten gesture categories including left over, right over, stop signal, move straight, right turn, left turn, right, left, lane left which are followed by lane right also exist as separate categories. This specialized dataset includes few images that were captured under different lighting conditions and ambient situations during gesture performance to enhance model generalization capabilities. The proposed hybrid architecture combines DenseNet121, due to its strength and robustness in discovering complex high-level patterns, with MobileNetV2 for its fast and efficient mid-level pattern recognition, all together combined with baseline CNN for its basic low-level feature analysis. The model merges different characteristics through its feature aggregation layer to maintain a balance between classification precision and system resource efficiency. During testing with the dataset, the hybrid architecture proved its suitability for real-time applications through achieving 98.7% accuracy at 1.03 s model processing time. The system works to address the challenges involved in traffic police gesture detection. The suggested system allows autonomous vehicles to automatically recognize and appropriately respond to driver gestures during real-life driving scenarios. The research represents a breakthrough in gesture recognition that enables intelligent transportation because it enhances autonomous driving system operational safety and effectiveness.