Real-Time Object Detection for User-Defined Few-Class Targets on Edge Devices

Object detection models have achieved remarkable performance in recent years, primarily through the use of large-scale datasets and computationally intensive architectures. However, in many real-world applications, only a limited number of object classes are relevant to the task at hand. This motivates the need for efficient, targeted detection methods that can operate under constrained environments such as edge devices. In this paper, we propose a real-time object detection framework that enables user-defined few-class selection from a pretrained base detector, optimizing both computational efficiency and task-specific performance. Our approach leverages heuristic class selection and dynamic model adaptation to maintain high accuracy for selected target classes, while significantly reducing resource consumption. Extensive experiments on benchmark datasets and various edge hardware platforms demonstrate that our method achieves a favorable trade-off between accuracy and speed, making it well-suited for on-device applications such as autonomous driving and surveillance.