OpenTLD is an open source implementation of the TLD (Tracking-Learning-Detection) framework, designed for real-time 2D tracking of a single object in video sequences. Because it fuses tracking and detection, TLD can recover from occlusions, drift, or failures by using its detection mechanism to reacquire the object. In terms of usage, one typically initializes the tracker by providing a bounding box on the first frame, then calls a function like run_TLD to process a video and obtain bounding boxes over time. The system updates its internal models as frames are processed, and can re-detect the target when tracking fails. The algorithm’s performance is known to improve over time due to its online adaptation behavior. Because of its age and MATLAB dependencies, adopting it in modern C++ / real-time pipelines may require effort (e.g. rewriting or porting) or using more recent tracking libraries.
Features
- Track unknown object from given bounding box
- On-line adaptation / learning of appearance (P-N learning)
- Detector module to re-identify object when tracking fails
- Support for handling occlusion / reappearance
- MATLAB / MEX support and C++ builds
- Integration with OpenCV / cross-platform build via CMake