多重感測視角系統中非接觸式的模型化體積測量

Abstract

本論文藉由非接觸式的模型重構達成物體型態測量，以取代直接接觸物體所得到的體積測量。2018推出的Intel RealSense深度攝影機提供簡便的介面將接收到的資訊轉換為三維點雲模型，可進一步透過計算幾何原理估算模型的體積。然而，深度感光設備先天存在一定的信噪比（Signal-to-Noise Ratio, SNR），使得輸出模型雜訊與觀測距離成正比，無法在相隔較遠的條件下對目標物得出高精度的測量值；再者，有鑑於單一深度攝影機僅能得到場景中特定視角的深度資訊，在攝影機照不到的區域（如:背面）將形成死角，難以完整還原全域的資訊。為了解決上述問題，本研究以多重攝影機陣列取代單一鏡頭補償投影成像資訊量的不足，並提出自動化的流程，針對落在多重攝影機視角共同範圍內移動中的物體計算出即時的體型估計值。在本研究提出的流程中，舉凡時間同步、多重視角座標轉換、物體辨識去除背景以及合成點雲表面曲線平滑化等皆是必須克服的問題。實驗結果證實由本研究設計的流程所得出的模型體測值，不僅適用於測量物體靜止時，對於移動中的物體進行非接觸式體型測量時也有一定程度的準確性。

This thesis proposes a method to measure body volume indirectly based on non-contact object point cloud modeling, providing an alternative to the traditional approaches that require laborious configuration and tremendous efforts. Recently, the improvement on three-dimensional sensor techniques, such as depth cameras, features high resolutions along with stability. Researchers find it suitable to apply depth cameras to the body volume measurement of livestock, which keep it from being disturbed during the measurement. Since the projected view of a single depth camera has limit surface information, in order to produce a complete 3D model that contains each side of the object for further measuring, we design a procedure that can fast and efficiently calibrated and align 3D point cloud models from multiple views. The measurement can thus be fulfilled on the synthesized point cloud models. In addition, the stereoscopic sensors are prone to be noisy due to the systematic errors. The noises should be removed to achieve a high precision of modeling. Therefore, we conduct quantitative analysis regarding to the noises. It turns out that the noises shall be eliminated with the appropriate filter chosen according to the analysis. The process of how to produce a comparable model of the object will be investigated thoroughly in our thesis. Once the model is built, we find out the desired body values, including body length, body width, heart girth, and body height through computational geometry. We claim that the error rate between our non-contact measuring and the tradition methods is under a certain percentage, which demonstrate the practicability of our research approach.