TLD Vision is focused on real-time tracking of objects in videos. Our technology is based on 3 pillars.
Video contains incredible amounts of data. With 30 fps and HD1080 resolution, this amounts to 248 MB/second. Luckily, this data volume is heavily correlated both in space and time. The temporal correlation is exploited in our tracking pillar which follows object motion assuming that the object appearance does not change much between neighboring frames. The trajectory of the target can be represented as a curve in the video volume.
Physical objects can appear at one location only. So given the trajectory of the target, it is safe to say where the object is and where is not. In other words, we can label the video volume by positive and negative labels. These positive and negative labels are fed into a classifier which is trained in real-time to recognize the tracked object even if there are similar targets in the scene. This classifier captures memory about the object appearance observed so far.
The memory captured by the classifier is used to localize patterns observed in the past. This is achieved by the detection pillar, which scans every location in the video stream and identifies locations which have high classifier response. This response is then integrated with the tracking response. Detector has the effect to re-detect target after occlusion or stabilize the tracking in background clutter.
TLD combines tracking, learning, and detection in a compact, real-time process. This gives to our technology the ability to improve in runtime and still maintain robustness necessary for any practical tracking system.
TLD2.0 was developed in 2012 from scratch in C++11. Its key innovation is more efficient detection module and more robust learning strategy. TLD2.0 is significantly faster and allows tracking multiple targets simultaneously.
TLD2.2 is our latest version which further improves the speed of the tracker, offers a number of low-level optimizations and pre-defined tracking profiles. It is the first version which runs on embedded devices.
xTLD is our secret new tracker. It is using deep learning and efficient online adaptation. Planned release is middle of 2017.