To build robust and accurate traffic sign recognition systems in diverse and complex environments, it is necessary to use a specific hybrid dataset creation technique, which combines both real-world and simulated datasets. We present an algorithm for data generation using blending that allows for traffic sign embedding on realistic roadside backgrounds with consideration of a contextsensitive data generation technique suitable for the local roads of Bangladesh. In contrast to traditional geometry-based augmentation, our solution retains the environmental elements such as lighting, shadows, occlusion, and real-world sign placements observed by BRTA, thus considerably narrowing down the domain difference between the synthetically generated training data and the reallife application. Together with the hybrid deep learning model framework, our solution is shown to increase the generalization and robustness of the traffic sign recognition system. Our contribution to the area of intelligent traffic sign recognition involves the introduction of a context-aware blended data set creation method as well as the validation of the hybrid deep learning architectures that enable more effective traffic sign detection in real-world scenarios. The resulting dataset contains 18,750 images spanning 46 traffic sign categories, with 70% blending-based images and 30% real images. Three models YOLOv8, YOLOv8 with VGG19, and YOLOv8 with ResNet50 were evaluated under different training and testing setups, achieving F1 scores of 0.952, 0.947, and 0.957, and mAP@50–95 values of 0.660, 0.668, and 0.670, respectively. Among them, the YOLOv8-ResNet50 model achieved the highest performance. Our findings highlight the effectiveness of hybrid dataset generation in advancing traffic sign detection and recognition in practical environment.