| DC 欄位 |
值 |
語言 |
| DC.contributor | 資訊工程學系 | zh_TW |
| DC.creator | 楊鈺瀅 | zh_TW |
| DC.creator | Yu-Ying Yang | en_US |
| dc.date.accessioned | 2013-6-28T07:39:07Z | |
| dc.date.available | 2013-6-28T07:39:07Z | |
| dc.date.issued | 2013 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=100522085 | |
| dc.contributor.department | 資訊工程學系 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 考量到維護安全及秩序的目的，基於影像資訊發展的智慧型監控系統成為熱門的研究題目。過去的監控式攝影機需要人力操作，透過人為監控，判斷是否有異常狀況發生，才能進一步針對特定狀況作處理。經歷數十年來的發展，運用圖形識別及影像處理技術的智慧型監控系統已可取代傳統的人力操作，應用於不同的環境中。其中，基於流量監控或異常路況偵測等不同目的之智慧型監控系統，廣泛地被應用於各國的重要交通路網。多數的智慧型監控系統，著重於行車狀況的監控。本研究希望能藉由監視影像的分析，自動取得交通場景的資訊。由位置較高的攝影機，取得特定路段的影像。取這段期間內行經此路段的車輛，其規律的移動資訊作為分析的對象，進而獲得此路段的道路資訊。後續更可利用所取得的道路資訊，更新導航軟體的路網，或者政府機關的即時路況資訊等等。
研究中所使用到的監視影像，拍攝的位置較高，距路面較遠。在拍攝位置較遠的狀況下，所得到的監視影像不同於一般的車流監視影像，有以下特性：拍攝範圍較廣，可涵蓋多條路線；車體較小，且在畫面中呈現為長方形色塊，較無立體感；雜訊較多，如行人、飛鳥等生物在經過畫面時面積極小，並有規律性地移動，干擾軌跡擷取及後續分析。
本研究首先利用連續影像相減法，取得畫面中有移動的部分，再對這些區域擷取特徵，再進行特徵點的追蹤，取得軌跡。將追蹤所取得的軌跡作分群，畫面中的軌跡根據所在位置及方向資訊，被分入不同的群聚中。以軌跡位置與群集中心所算出的變異數大小作評估，希望能取得分布最為密集的分群結果。分別算出各條軌跡的主要前進方向，並針對各個群聚，將主方向不同的軌跡分割出來。分割後，依據各個群聚中軌跡取樣點位置的累計直方圖，推測車道的分隔線，再分割出不同車道的群聚。最後分析各個群聚周邊的相鄰群聚，依據相鄰群聚的位置和主方向的資訊，連結各個路線，最後得到畫面中的路向和路口資訊。 | zh_TW |
| dc.description.abstract | For traffic safety and efficiency, the surveillance system based on image analysis has become popular research project in recent years. Intelligent surveillance system using pattern recognition and image processing technologies have been developed and improved intensively in this decade. In the thesis, we propose a method to automatically segment the regions of interest in the surveillance videos via analysis of vehicle trajectories. The system aims at dealing with wide-range surveillance, which is an important complement to ground-plane surveillance. The traffic scenes in experimental videos are taken from high buildings. Therefore, the vehicles in the scene are not stereoscopic. The pedestrians are just like black spots, and there are many noises in the scenes.
First, we do inter-frame differencing and dilation to get the motion region. Then, we perform tracking using the ORB feature, HOG and the color histogram in the motion region. And the tracking results are stored as the trajectories. Afterwards, we use spectral method to automatically determine the number of clusters, and use K-means with modified Hausdorff distance to cluster the trajectories. Then the clusters are segmented with the orientations of trajectories. Assuming that most of vehicles move along the lanes, and lane changing seldom appears. We apply the point rotation and projection to get the valley in the distribution range of each cluster. The position of valley implies the position of lane. After the segmentation of lanes, we merge the overlapping cluster. Finally, to get the correct road map, we do connection with neighbor cluster and orientation. We conduct experiments with different challenging surveillance scenes to validate the proposed method. | en_US |
| DC.subject | 軌跡分群 | zh_TW |
| DC.subject | 豪斯多夫距離 | zh_TW |
| DC.subject | 路面分析 | zh_TW |
| DC.subject | 感興趣區域切割 | zh_TW |
| DC.subject | 智慧型監控系統 | zh_TW |
| DC.subject | Trajectory Clustering | en_US |
| DC.subject | Modified Hausdorff Distance | en_US |
| DC.subject | Route Analysis | en_US |
| DC.subject | Region of Interest Segmentation | en_US |
| DC.subject | Intelligent Surveillance System | en_US |
| DC.title | 自動感興趣區域切割及遠距交通影像中的軌跡分析 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.title | Automatic Region of Interest Segmentation and Trajectory analysis in Far Distance Traffic Surveillance | en_US |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |