Real-Time Pedestrian and School-Zone Sign Detection for Autonomous Driving Using Spatial Attention-Enhanced Faster R-CNN

Abstract

Pedestrian detection and School-zone sign recognition are crucial aspect of advanced driver assistance systems and intelligent transportation systems. Consistent perception improves traffic safety and enhances the reliability of Advanced Driver Assistance Systems (ADAS). This research proposes a Spatial Attention-enhanced Faster R-CNN framework for pedestrian and school-zone sign detection. The combination work effectively in dense environments with varying light levels, as well as being constrained by partial occlusion. Experiments on the WIDER Person dataset compare two backbone designs, MobileNetV3 and ResNet50, which both use Spatial Attention and Faster R CNN. While MobileNetV3 is used as a lightweight baseline testing and for an ablation study and edge deployment, ResNet50 is selected as the primary backbone architecture due to its superior feature extraction capability. A Spatial Attention enhanced ResNet50 model was trained using data obtained from the Roboflow repository consisting of images of school-zone signs, captured in real-world conditions. the Spatial Attention Module significantly contributed to the ability of the model to perform feature refinement through concentrating on the most applicable spatial regions, thereby improving the spatial localization of small and partially occluded objects. All models are implemented and evaluated on the NVIDIA Jetson Orin Nano platform to investigate their suitability for embedded ADAS deployment. Hardware metrics including latency, throughput, GPU utilization, and power consumption are evaluated to assess embedded deployment feasibility using the baseline PyTorch implementation without deployment-oriented optimization techniques such as TensorRT acceleration, pruning, or quantization. Experimental results demonstrate stable and accurate detection performance, establishing a practical baseline for embedded online perception while providing a foundation for future deployment optimization in ADAS applications.