Research Of Variational Optical Flow Computation On Heterogeneous Platforms

Optical flow method(i.e., the calculation method of optical flow field) in pattern recognition, computer vision and image processing, and other fields has a very important basic position. Because of optical flow field is an object moves in the space field in image plane corresponding to reflect, by calculating the light in the camera movement and the flow field can be in the information of the scene under the condition of unknown detect motion object. So a large number of applications such as video surveillance, video retrieval, medical imaging, automatic navigation, 3D scene reconstruction and so on all need to use the optical flow method.In different optical flow method represented by the HS(Horn and Schunck) variational optical flow method, because it can calculate more dense optical flow information and the recent progress in dealing with a larger displacement,it get high attention by the researchers. But variationa optical flow method was applied to the actual system, the biggest difficulty is its computation complexity,and real-time performance. For example, In the configuration frequency 3.4 GHz Intel Core i3-3240 CPU and 4G memory of PC running the HS dense optical flow method, When the image resolution is 640 * 480, the average time need more than 2 seconds per frame, So it is an urgent need to improve the computing speed.As the power consumption is more and more become the important factor of computing system must be considered, computing performance can no longer rely on the improvement of working frequency to ascend, but more to improve parallel processing ability of the system. And research shows that further, heterogeneous computing platform has an advantage in computational efficiency with different characteristics calculation unit, such as GPU.Therefore, this article mainly research variationa optical flow optimization calculation methods in different heterogeneous computing platforms such as the CPU + GPU and CPU + Tilera and so on. First studied the calculation characteristics of variationa flow optical method, analyzes the bottleneck and different calculation module of parallelism, then separately in two different parallel computing platfomr. The experimental results show that, two kinds of heterogeneous platform both in energy consumption and speed is better than the CPU, even GPU is better than CPU dozens of times. When the resolution of the image is 640 * 480, GPU processing two frame takes 24 ms 83 times faster than the CPU, energy consumption is only about 1/21 of the CPU. Tilera takes 0.8 s 2.5 times faster than the CPU, energy consumption is only about 1/6 of the CPU. To a certain extent the results show the optical flow is more suitable for data parallel way. Although CPU + GPU mode it needs less energy to run the optical flow method, but due to its high power led to higher energy consumption under the idle state. In addition, the parallel computing method on the two platforms with the increase of the cell has a good scalability. This topic research results to reveal with low power consumption, high real-time requirement of the calculation method of optical flow method which is more suitable for the embedded environment.