基於雷射測距儀之多移動物體追蹤的資料分群與連結標示系統

Abstract

為了要讓機器人了解周遭的動態環境，成功地追蹤移動物體是不可缺少的必要條件。在過去，學者們陸續提出了許多移動物體追蹤的演算法。然而到目前為止，卻沒有足夠的、經標示過的真實資料可以用來評估、比較各個演算法的效能。面對這樣的瓶頸，我們針對基於雷射測距儀的移動物體追蹤演算法，設計了雷射資料分群與連結的標示系統。使用者可以藉由這系統標示各個移動物體在二維空間中實際代表的雷射點群，進而與演算法計算的結果相互比較。由於資料標示的工作繁重且力求精確，我們的設計著重於提升使用者標示的精準度以及減少使用者的工作負荷。經由實驗，不僅證實了我們的標示系統能確實提升使用者的標示成果，也提供了標示過的交通路口資料，可以用來評估追蹤演算法在高度動態都市環境中的表現。

Scene understanding is one of the most important foundations for a mobile robot to operate in human-habited environments. As the real environments are typically dynamic, moving object tracking becomes an inescapable problem. While the tracking algorithm becomes more and more elaborate, however, its performance in real world still can not be guaranteed. The major reason is that so far we do not have enough real data with ground-truth to evaluate and analysis the state of the art tracking algorithms . In this thesis, we explore the laser-based moving object tracking problem and propose an annotation system that allows the user to annotate the ground-truth of segmentation and data association with 2D laser measurements. As the annotating task is difficult and tedious, the system is designed to achieve higher accuracy and reduce the task loading in the annotation process. To prove the usefulness of our system, real data sequences are collected and annotated by multiple users in our experiments. The results shows that the annotation performance varies but the system keeps helpful across different users. In particular, the V-measure reaches to 0.995 bits and the false positive rate and the false negative rate are reduced to 0.341% and 1.239%. At last, the ground-truth data is also generated by validating the annotated data carefully and repeatedly.