High-frame-rate region-based visual tracking on CPU: an implementation perspective
High-frame-rate region-based visual tracking on CPU: an implementation perspective
复制标题
CPU 上基于高帧率区域的视觉跟踪：实施视角
DOI：
10.1109/sii.2016.7844058
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
Koichi Hashimoto
中科院分区：
文献类型：
--
作者：
Shingo Kagami;Koichi Hashimoto
本文描述了高效二阶最小化方法的快速实现，基于区域的平面目标的视觉跟踪的多核CPU上的SIMD指令，而不依赖于GPU。我们表明，仔细的实施考虑导致高性能，即使只有标量指令在一个内核上，并充分利用并行实现了超过3500 fps的模板与64 × 64像素和1200 fps的128 × 128像素的英特尔酷睿i7- 4600 U处理器。通过450 fps的USB 3.0实时视觉跟踪实验，验证了该算法在高帧率跟踪应用中的有效性。
This paper describes fast implementations of Efficient Second-order Minimization method for region-based visual tracking of a planar target on a multi-core CPU with SIMD instructions, without relying on GPUs. We show that careful implementation considerations lead to high performance even only with scalar instructions on a single core, and full utilization of the parallelism achieves a frame rate of more than 3500 fps for a template with 64 × 64 pixels and 1200 fps for 128 × 128 pixels on Intel Core i7-4600U processor. Real-time visual tracking experiments with 450-fps USB 3.0 shows its effectiveness in high-frame-rate tracking applications.