- 期刊
PAC雙核心系統晶片平台上的即時物件偵測系統
Implementation of Real-Time Object Detection System on PAC Duo Platform
摘要
隨著文明科技發展及生活水準的提高，電腦視覺應用如遠端照護、監控系統等，已漸漸在人們生活佔有一席之地。物件偵測作為電腦視覺應用之前期基本程序，文獻上已存在許多設計縝密的演算法，且其多以資源充沛的桌上型電腦作為實現平台；然而基於成本、方便性等多方考量，各應用最終仍需和嵌入式系統作一整合。本文以工研院系統晶片科技中心自行研發的PAC雙核心系統晶片(PAC Duo SoC)為實作平台，利用PAC數位訊號處理核心(PAC DSP)支援超長指令集架構的特性，搭配演算法的平行最佳化設計，並輔以組合語言的程式撰寫技巧，進而將效能做最有效發揮。實作結果得知，針對解析度為VGA(640×480)的輸入影像，本系統以150 MHz的工作頻率，即可在保有偵測正確性的前提下，達即時偵測之處理效能（每秒30張畫面）。
並列摘要
With the explosive growth in technology, computer vision applications such as telecare, surveillance system have been an important topic in our real life. In the past, lots of computer vision applications and algorithms are implemented based on PC without considering the penalty of space and cost. Embedded system is a promising platform for the development of vision related applications. In this paper, we will propose a robust object detection system based on the PAC Duo Platform developed by ITRI/STC. The performance is improved by using the optimized parallelism design and adoption of assembly language to enhance the utilization of hardware capability. Experimental result shows that the proposed system can achieve real-time capability (30 frames per second) of object detection with VGA resolution (640×480) at the DSP frequency of 150 MHz.