Adaptive Input Sampling: A Novel Approach for Efficient Object Detection in High Resolution Traffic Monitoring Images

Real-time traffic monitoring based on artificial intelligence (AI) models is essential for efficient transportation management, but it faces significant computational challenges, particularly when processing high-resolution video footage. Modern traffic monitoring systems typically rely on AI-based object detection and tracking models, which require processing every frame of the video to detect and track vehicles. However, even the most efficient object detection models are computationally expensive, making real-time processing of high-resolution videos impractical. To address this challenge, we introduce the Adaptive Input Sampler (AIS), a preprocessing module designed to selectively activate the object detector only when vehicles are present in the scene. The AIS dynamically targets specific areas of the frame where vehicles are detected, sampling these regions for downstream processing. This approach significantly reduces the computational burden on the object detector while preserving the original pixel details, enabling substantial computational savings—especially in scenarios with few or no vehicles in the monitored scene. Our experimental results demonstrate that the AIS achieves 96.40% traffic counting accuracy and 95.60% vehicle classification accuracy, while reducing the original input data size by up to 79.00%. By optimizing the volume of input data, the AIS improves computational efficiency and accelerates inference results, making it a promising solution for real-time traffic monitoring applications. The AIS represents a significant step forward in addressing the computational challenges of high-resolution video processing, offering a balance between accuracy, efficiency, and scalability for modern transportation systems.