Accurate Real-time Detection of Vehicles and Pedestrians on Edge Device with Scaled-YOLOv4 and TOPST AI

Advancements in deep learning-based object detection offer high accuracy, but often require computationally expensive models. In driver assistance systems (DASs) for autonomous vehicles, real-time performance and low power consumption are crucial. This study investigates the optimization and deployment of the lightweight Scaled-YOLOv4 object detection model on a power-efficient embedded platform equipped with a Neural Processing Unit (NPU). By leveraging the TOPST AI edge device, we aim to achieve a balance between performance and efficiency. Experiments on K-SoC traffic dataset demonstrate the effectiveness of this approach in delivering real-time, high-accuracy vehicle and pedestrian detection on resource-constrained edge NPU systems.