To speed up VR controller pose estimation, a HoG detector (Histogram-of-Oriented-Gradients) was developed to detect ROI of VR controller from a single image. Random forest classifier was trained using HoG features and moments of LED observations inside a test ROI. Random forest training was done using collection of controller images along with LEDs in Python environment. Negative samples were also collected using random background image patches. Python detector was exported into linkable C++ code. HoG detector proved to be fast on CPU and avoided GPU DNN frameworks completely.
HoG detector tested using real camera (pin-hole compatible).
HoG detector trained and tested using simulated fisheye video.