可被電腦視覺模組驅動的粒子系統之開發

Abstract

近年來以電腦視覺為基礎的互動裝置越來越受歡迎，對於能給予使用者有趣互動體驗以及豐富視覺效果的應用軟體需求也因應而生。本研究開發出一個可被電腦視覺模組驅動的粒子系統，能輕易地被整合到各種電腦視覺的應用程式，提供多樣化的視覺特效和互動。另一方面，由於電腦視覺的輸出結果通常帶有一定程度的不確定性，粒子系統也為此提供了一個視覺化的方法，用粒子的分佈與擾動將這個模糊地帶顯現出來。本篇論文中，我們舉出三個整合範例，包括光學式多重觸控桌面系統i-m-Top，多重攝影機監控轉場系統，以及智慧型太極拳輔助教學系統，三種完全不同的應用程式皆可與我們提出的粒子系統輕易整合。為了達到更好的執行效率，我們在粒子系統的實作上採用新一代的繪圖處理器(GPU)來加速執行，此舉不但大幅擴張了粒子系統中能即時運算的粒子數量，同時也將中央處理器(CPU)的運算資源留給電腦視覺模組。另外我們也提出一個粒子間相互碰撞偵測的方法，利用統計的方式估算出粒子在空間中的密度分佈，並依此估計粒子所受到的碰撞效果。此方法讓粒子產生類似布朗運動的行為模式，並能快速地以繪圖處理器運算出結果。

Vision based interaction devices are getting more and more popular in recent years. Applications for these devices that can give users interesting interaction experiences and rich visual appearances are also on demand. In our work, we develop a particle system that can be driven by computer vision modules. It is easy to be integrated into other computer vision applications and provides varieties of visual effects and interactions to the users. On the other hand, results of computer vision process usually have some degrees of uncertainty. Our particle system also supplies a new way to visualize the fuzzy region by the motion of particles. In this paper, we introduce three integration examples, including a vision-based multi-touch tabletop system called i-m-Top, a multi-camera surveillance system with view transition, and an intelligent computer-aided Tai Chi Chuan learning system. All these three applications can be easily integrated with our particle system. To increase the system performance, we utilize the latest generation of graphic process unit (GPU) to implement our particle system. This implementation not only enlarges the number of particles we can simulate in real time, but also reserves the CPU computation power for the computer vision process. In addition, we propose an inter-particle collision detection method that estimates the spatial density distribution of particles by statistic method. Inter-particle collision responses are then computed according to the density distribution. This method makes particles have behaviors similar to the Brownian motion and can be done by GPU very efficiently.