摘要
本論文中，我們提出了一種改良式粒子群演算法，名為單維搜索分工式粒子群演算法(Particle swarm optimization with one dimension multi-modes, ODMPSO)，並應用ODMPSO演算法於影像追蹤系統。標準粒子群演算法中每一個體使用相同的移動方程式，而本文所提的方法，能使粒子根據其位置狀態選擇其移動方程式進行位置的更新。在粒子群最佳化初期，透過特殊的單維搜索機制，讓粒子可以更有效的從局部探索開始，逐漸演化，等到集中收斂至一階段後，再使用分群機制，根據其粒子位置分別置入數個子群中，在子群中的粒子根據其對應的四種模式進行速度更新，以求迅速的把其它個體帶往全域最佳解。本文並利用ODMPSO演算法提升影像追蹤系統上的效能，獲取更好的辨識率與更快的迭代速度。由於我們得知在傳統的高斯混合模型背景相減法(Gaussian mixture model background subtraction)裡，所使用的迭代方式是使用期望值最大化演算法(Expectation Maximization , EM)，而此方法在進行迭代時，緩慢的收斂速度，往往影響了即時的影像辨識系統之實用性，所以本文採取ODMPSO演算法來提升收斂速度，以防止耗費大量的運算，減少系統的運算的複雜度。從實驗結果證實，所提出的ODMPSO可以得到較佳的平均值(Mean)、標準差(Standard deviation)與辨識率，並且能大幅地提升系統的收斂速度，所以證實所提出的演算法的確能有效地增進影像追蹤系統的實用性。
並列摘要
In this thesis, we propose a modified particle swarm optimization algorithm which is called particle swarm optimization with one dimension multi-modes (ODMPSO). The proposed ODMPSO which is different from standard PSO algorithm is moving functions. In ODMPSO method, the particles can be adaptively searched by their environment. There are five modes in ODMPSO method. Each mode has its own specific optimizations. Finally, these modes makes the particles more easily and quickly find the results. Afterwards, we propose a Gaussian mixture model based on ODMPSO (GMM-ODMPSO) method in a visual tracking system. The GMM-ODMPSO method will accelerate the convergence rate of creating the GMM background model and the system also improves the detection of moving targets. The experimental results show that the proposed GMM background model obtains better recognition rate. As seen in the experiments, the GMM-ODMPSO method is a 48% improvement over the computing time, 88% over the convergence rate, and the recognition rate is almost the same as the traditional GMM background model. In the results, we can see our proposed method is more effective.