- 學位論文
基於共享分佈式粒子濾波器之多相機目標追蹤
Multiple Camera Tracking Based on Distributed Particle Filter with Consensus Algorithm
摘要
目標追蹤與估測的技術應用從早期的軍事衛星、飛航與導彈系統到現代國安偵防就被廣泛的研究，而隨著時代推進，現代化的科技與各種友善便利的開發環境，這類的軍事科技也開始在民間上大量的被使用，不論是生態研究、救災防護甚至到最近新興的無人自動駕駛技術等都是具體實現應用的例子。 本篇論文以分別放在三個不同位置相機並將畫面中心對準地板上的世界座標原點後即時傳遞影像，畫面經局部二值模式辨識目標成功後交由連續自適應均值漂移保持追蹤並降低所需要的計算量，再以卡爾曼濾波器輔助估測目標移動狀態來降低目標丟失率，相機座標經由歐拉角轉換成世界座標後以分佈式粒子濾波器依權重來作目標移動估測修正，最後三個裝置彼此共享資料再經由共識演算法決定出目標最後的位置。 藉由多裝置的共識演算法可以實現即使當三個裝置中有目標遭到遮蔽時，其他的裝置也可以藉由共享情報與狀態估測來繼續保持目標位置並協助重新追蹤。其中可藉由增加粒子濾波器的數量來提升預估目標狀態的穩定度，但也會因此消耗更多計算時間與資源，而調整共識演算因數可以分別觀察在目標有丟失以及無丟失的情況下共識資料的擾亂程度與精準度，再依狀況決定最佳的參數，最後研究模擬不同二維迴避模型的差異對於三裝置的共識率分析與探討，以此來找出高效能與低耗時的最佳解。
並列摘要
The techniques of object detection and tracking are studied extensively for various applications, ranging from early military satellite, navigation system to modern national security. As time progresses, advanced technology and approachable development environment allow us to more easily apply these military technologies to social purposes such as ecological research, disaster relief, even to new technologies of self-driving cars which has attracted extensive attention worldwide. In this thesis, we discuss multi-agents image tracking, which uses local binary patterns, Camshift and Kalman filter, and multi-agent tracking, which uses particle filter, Poisson process model and consensus algorithm. For image detection and tracking, local binary patterns (LBP) is the first technique we apply to target recognition, which is more suitable for simple objects such as an archery target. Further, the Camshift scaled search algorithm and Kalman filter-based recursive estimation can maintain track and reduce target lose. To achieve multi-agent tracking, multiple particle filter which have an individual weight of moving state and the Poisson process are employed to determine object location. Each part filter or agent models target maneuver by a Poisson process and consensus algorithm. By collecting all agents' data and modifying the value of consensus coefficient, the final result of object location will be produced. Experimental implementation has also been conducted. Three individual cameras are deployed in the different locations, aiming centered at the origin of a world coordinate and capture images in real time continuously. By using LBP recognition, when the object in the video is recognized and detected successfully, Camshift takes over and keeps track with lower computation and faster response. A Kalman filter is used as auxiliary to estimate state and reduce the rate of target lose. Rotating camera coordinates to world coordinates via Euler angles rotation matrix, and implementing distributed particle filter, accurate estimate of the target's world position is produced. Finally, the consensus algorithm is a computationally efficient method to fuse the estimates of the agents. makes the final result of the object's location by computing data, which received from three individual agents sharing with each other.
參考文獻
延伸閱讀
- Lai, C. W. (2006). 具遮蔽處理之貝氏濾波法實現主動相機平台之多目標物影像追蹤 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU.2006.01834
- 鄭宇珊（2005）。複雜背景下多相機之多目標物影像追蹤系統〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU.2005.00781
- Song, C. J. (2007). 循序權重取樣粒子濾波法實現全方位相機之多目標物影像追蹤 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU.2007.02403
- 范夢葳（2012）。僅使用角度追蹤目標物的交互性多模型延展卡爾曼與粒子濾波器〔碩士論文，國立中央大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0031-1903201314455595
- Huang, K. Y. (2009). Object Tracking of Multiple Cameras [master's thesis, National Taipei Uinversity]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0023-2108200917500400