基於相機參數之相機移動補償;Camera motion compensation based on camera parameter

NCU Institutional Repository > 資訊電機學院 > 軟體工程研究所 > 博碩士論文 > Item 987654321/95257

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/95257

题名:	基於相機參數之相機移動補償;Camera motion compensation based on camera parameter
作者:	詹翔淵;Chan, Hsiang-Yuan
贡献者:	軟體工程研究所
关键词:	物件追蹤;相機參數;相機移動補償;Object tracking;Camera parameter;Camera motion compensation
日期:	2024-07-15
上传时间:	2024-10-09 16:36:03 (UTC+8)
出版者:	國立中央大學
摘要:	物件追蹤，旨在連續影像序列中檢測出目標位置，並且預測其運動軌跡，逐幀更新目標的位置，要完成這項任務，現今多採用兩種特徵將不同幀之間的相同目標關聯起來，其一是透過目標的外觀特徵透過匹配前後幀偵測框中的像素特徵來得到目標當前的位置，其二是通過目標的運動特徵，利用 Kalman filter 等動態系統模型，藉由目標的歷史位置來預測目標的當前位置，得到兩種特徵後，將兩種特徵按照不同的情況，對其採用不同的權重比例計算預測框及檢測框之間的距離成本，最後透過匈牙利演算法，將成本最低的預測框及檢測框進行匹配並賦予其相應的 ID 完成追蹤。在動態相機的情況下，其中的運動特徵如果只透過傳統的 Kalman filter 來進行預測可能會導致偏差，原因是當相機移動時，影像的像素會根據相機的移動方向產生反向的偏移，而這個偏移是 Kalman filter 無法自行估計的，這將造成預測出來的目標位置產生偏差進而造成目標丟失。為了解決這個問題，2022 年，BoT-SORT 模型透過 CMC 相機移動補償 (Camera Motion Compensation) 來處理這個問題，他們利用稀疏光流來預測特徵點在幀與幀之間的移動，進而計算出相機移動造成的誤差，並提出了一套方法利用計算出的誤差對 Kalman filter 的預測進行修正。這一方法使得動態相機下的追蹤效能得到了顯著的進步。然而，在這其中仍然存在改良的空間，用來計算偏差的稀疏光流，其帶來的計算量仍是不可忽視的，並且由於是基於影像的特徵點進行預測的關係，影像的品質在一定程度上也影響了預測的準確度。在本文中，我們提出了一種透過相機參數來計算相機移動誤差，並透過相機移動補償機制對 Kalman filter 進行校正的方法，該方法透過針孔相機模型的原理轉換3D 空間的相機移動與 2D 空間的影像偏差，不僅降低了稀疏光流造成的龐大計算量，只需要低廉的計算成本，並且可以避免依賴影像的特徵點，同時，我們透過目標追蹤相關的論文得到啟發，在關聯階段將預測框以及偵測框根據目標之間所處的位置進行擴張，減少因為微小誤差而導致的追蹤丟失，這使得我們的追蹤模型可以達到更快速、更穩定的效能。;Object tracking aims to detect the position of a target in a continuous image sequence,predict its movement trajectory, and update the position of the target frame by frame.To accomplish this task, two features are currently used to associate the same target between different frames. One is to obtain the current position of the target through theappearance characteristics of the target by matching the pixel features in the detectionframe of the previous and next frames. The second is to use the motion characteristicsof the target and use dynamic system models such as Kalman filter to use the historyof the target. position to predict the current position of the target. After obtaining thetwo features, the two features are used according to different situations, using differentweight ratios to calculate the distance cost between the prediction frame and the detectionframe. Finally, through the Hungarian algorithm, The lowest-cost prediction frame anddetection frame are matched and assigned corresponding IDs to complete tracking.In the case of a dynamic camera, if the motion features are predicted only through thetraditional Kalman filter, it may lead to bias. The reason is that when the camera moves,the pixels of the image will be shifted in the opposite direction according to the movingdirection of the camera. This offset cannot be estimated by the Kalman filter by itself,which will cause the predicted target position to deviate and cause the target to be lost.In order to solve this problem, in 2022, the BoT-SORT model dealt with this problemthrough CMC camera motion compensation (Camera Motion Compensation). They usedsparse optical flow to predict the movement of feature points between frames, and thencalculated the camera Errors caused by movement, and a set of methods are proposedto use the calculated errors to correct the predictions of the Kalman filter. This methodsignificantly improves the tracking performance under dynamic cameras. However, thereis still room for improvement. The amount of calculation caused by the sparse opticalflow used to calculate the deviation cannot be ignored, and since prediction is based onthe feature points of the image, the quality of the image is limited to a certain extent.The degree also affects the accuracy of prediction.In this article, we propose a method to calculate the camera movement error throughcamera parameters and correct the Kalman filter through the camera movement compensation mechanism. This method converts the camera movement in the 3D space throughthe principle of the pinhole camera model. The image deviation from the 2D space notonly reduces the huge amount of calculation caused by sparse optical flow, but also requires low calculation cost and avoids relying on the feature points of the image. At thesame time, we were inspired by papers related to target tracking and related In this stage,the prediction frame and the detection frame are expanded according to the position between the targets to reduce tracking loss due to small errors, which allows our trackingmodel to achieve faster and more stable performance.
显示于类别:	[軟體工程研究所 ] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	34	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....