使用OpenCL加速AdaBoost行人偵測器之實作探討

Study and Implementation of Accelerated AdaBoost Pedestrian Detector on OpenCL

邱耕毫(Keng-Hao Chiu)

指導教授 ： 陳鵬升 ； 郭峻因

國立中正大學/工學院/資訊工程學系/碩士(2015年)

摘要

物件偵測是電腦視覺領域中相當重要的議題，如人臉偵測、行人偵測等，都是其重要的應用，而高偵測率往往在偵測速度上有所犧牲，其中一個解決的辦法是利用增加硬體去加速，但無法針對演算法去調整。
因此，本論文提出異質多核心平台 (CPU+GPU)協同運算來進行物件偵測的加速。在文中探討的物件偵測是利用AdaBosst與矩形特徵訓練分類器，主要針對佔了偵測大部分時間的「特徵比對」來進行加速，如以純CPU計算，還有可加速的空間。根據AdaBoost演算法移植到GPU上執行，會有偵測視窗工作量與尺度工作量分配的問題。本論文提出尺度工作平行化與工作量分配演算法來優化CPU與GPU協同運算之工作分配。
我們以CPU與GPU之架構，並運用其運算資源進行平行加速運算，由實驗結果得知在AMD A8-3850和ATI FirePro V8800的平台下，在處理720x480的影像中，可以達到fps 3.81與 Speedup 24.4；而在經過設定感興趣的區域後處理360x240的影像，可以達到fps 16.67與Speedup 20.08。

關鍵字

行人偵測

並列摘要

Object detection is an important research topic in computer vision. Many appli-cations (such as face detection, pedestrian detection) can benefit from this technique. However, high detection rates tend to be at the expense of the detection speed, due to high computational complexity. One of the solutions is to use specific hardware to improve the performance, but it is lack of the flexibility if we need to revise the algo-rithm.
In this thesis, we proposed a parallel algorithm for AdaBoost pedestrian detector, which is suitable to be implemented on a heterogeneous platform consisting of multi-ple CPU and GPU cores. AdaBoost classification with Haar-like feature is used in the proposed algorithm for the pedestrian detection. The most time-consuming part in the algorithm is feature calculation, which occupies over 98% of the computation. Thus, we accelerate the feature calculation by using both CPU and GPU computing re-sources. The strategies of scale-load parallelization and workload distribution are proposed to optimize CPU and GPU cooperative computing. The proposed algorithm was implemented using OpenCL for the evaluation. The experimental results show that the system can achieve 3.81 fps at 720x480 and 16.67 fps at 360x240 in the AMD A8-3850 and ATI FirePro V8800 platform.

並列關鍵字

pedestrian detection ； OpenCL

參考文獻

[6] Daniel Hefenbrock, Jason Oberg, Nhat Tan Nguyen Thanh, Ryan Kastner and Scott B. Baden, “Accelerating Viola-Jones Face Detection to FPGA-level using GPUs,” Proc. IEEE Annual International Symposium on Field-Programmable Custom Computing Machine (FCCM), pp.11-18, 2010.

[18] David Oro, Carles Fern´andez, Javier Rodr´ıguez Saeta, Xavier Martorell, Javier Hernando “Real-time GPU-based Face Detection in HD Video Sequences,” IEEE International Conference on Computer Vision Workshops, 2011.

[1] P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” Proc. IEEE Computer Society Conference on Computer Vision and Pat-tern Recognition, vol.1, pp.511–518, 2001.

[4] Henry A. Rowley, Shumeet Baluja, and Takeo Kanade, “Neural Network-Based Face Detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, No. 1, January 1998.

[8] Dmitry Pertsau, Andrey Uvarov, “Face Detection Algorithm Using Haar-Like Fea-ture for GPU Architecture,” Proc. IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems, 2013.