An Improved TLD Tracking Algorithm with RPCA‑Based Occlusion Handling and a Simplified Nearest Neighbor Detection Model

Abstract

The Tracking‑Learning‑Detection (TLD) framework is a real‑time long‑term tracking system comprising three components: tracking, learning, and detection. However, tracking performance is inevitably degraded by severe illumination changes, occlusion, and rotation, which frequently lead to tracking failure. To address these challenges, this paper proposes an improved TLD algorithm that integrates Robust Principal Component Analysis (RPCA) into the detection stage to separate low‑rank components from the image, thereby effectively suppressing background occlusion. In addition, the detection model is simplified by replacing the cascaded classifier with a Nearest Neighbor Classifier (NNC), which enhances both efficiency and effectiveness. By combining RPCA‑based occlusion handling with a streamlined detection model, the proposed algorithm achieves more robust and efficient dynamic tracking in long‑term scenarios..