Robust Low Complexity Feature Tracking using CUDA

In this paper, we propose a real-time video processing implementation of a Robust Low Complexity Feature Tracking (RLCT) algorithm on GPU (Graphics Processing Unit) using the CUDA (Compute Unified Device Architecture) paradigm. The RLCT outperforms state-of-the-art implementations of pyramidal KLT (Kanade-Lucas-Tomasi) on GPU by removing the overhead of the image pyramid construction, by predicting the initial tracking location for faster convergence and terminating the tracking once convergence is reached instead of executing for a fixed number of iterations. To track 1000 feature points on images of size 960×960, RLCT-CUDA implementation running on a GeForce 280 GTX GPU is ~25 times faster than RLCT on CPU and ~236 times faster than the original pyramidal KLT tracking algorithm on Intel Core 2 Duo 2.66 GHz with 2GB RAM CPU.