Investigating the Parallel Components of TLD Algorithm Using OpenCL Computation Framework

摘要

Long term object tracking is becoming more popular with the introduction of the Tracking-Learning-Detection TLD algorithm, and yet it has not been fully optimized to operate in scalable environments. It is essential to address some sections of the algorithm in terms of intense computations in order to cope the real-time requirements and boost the overall performance of object tracking. In this study, the core components of the algorithm that slow down the operation were investigated and implemented in parallel computational environments such as Multicore-CPUs and GPUs (graphics processing unit) with the use of OpenCL framework. Such implementations make it applicable for larger video inputs or higher frame-rates. The model then can be expanded to process multiple inputs simultaneously, and that parallelism brought speed up to the existing implementation. The implementation kernels are RGB to Gray, Sobel Filter and Variance Filter, and their performance evaluated similarly using different image sizes and implemented on different devices. According to the experimental results, for relatively small inputs the speed up for kernels is minimal, but it scales very nicely for large inputs. Speed ups are obtained as 2X for RGB to Gray conversion, 56.25X for Sobel Filter and 54.33X for Variance Filter.