This paper presents a novel automatic multiple moving target detection and tracking framework that executes in real-time with enhanced accuracy and is suitable for UAV imagery. The framework is deployed for on-board processing and tested over datasets collected by our UAV system. The framework is based on image feature processing and projective geometry and is carried out on the following stages. First, FAST corners are detected and matched, and then outlier features are computed with least median square estimation. Moving targets are subsequently detected by using a density-based spatial clustering algorithm. Detected targets' states are estimated using Kalman filter, while an overlap-rate-based data association mechanism followed by tracking persistency check are used to discriminate between true moving targets and false detections. The proposed framework doesn't involve explicit application of image transformations to detect potential targets resulting in enhanced computational time and reduction of registration errors. Furthermore, the selective template update mechanism that's based on the data association decision ensures sustaining a representative target template. Also, using BRIEF descriptors for target localization enhances framework robustness and significantly improves the overall tracking precision. Quantitative results are carried out on real-world UAV video sequences collected by our UAV system and on publicly available DARPA datasets. The experiments prove the robustness of the proposed framework for practical UAV target detection and tracking applications.
展开▼
