Enhanced TLD-based video object-tracking implementation tested on embedded platforms

Abstract

Object-tracking algorithms on embedded platforms are very important in many civilian and military applications. The Tracking–Learning–Detection (TLD) algorithm is considered one of the state-of-the-art online long-term object-tracking algorithms. The performance of running such computationally intensive algorithms on embedded platforms with limited computing resources is a challenge. This work proposes an enhanced TLD implementation, specifically designed for, and tested on, embedded platforms. In this new implementation, an extra-stage has been added to the TLD detector cascade, called a Region filter. This filter dynamically identifies the candidate region for the tracked object. Further, the two independent tracker and detector TLD components, and the two independent Forward–Backward (FB) and Normalized Cross Correlation (NCC) error measures in the tracker have been parallelized. Still further, the computations of Image Integral in the detector and the NCC in both the tracker and the detector have been optimized using a single instruction multiple data (SIMD) architecture. We evaluate our proposed implementation on the Apalis T30 embedded platform, using the same video sequences that the original TLD is evaluated on. Our results show that our enhanced implementation outperforms the baseline with an average speedup of 3.7 × in the total number of frames per second (fps), while achieving an average 91% of the Precision and 86.6% of the Recall metrics, across all sequences. Further, our enhanced implementation achieves an average speedup of 4.52 × and 1.86 × in the detector and tracker execution times, respectively. Moreover, it results in an average 66.3% energy saving, as compared to the original implementation.