智慧型運輸系統中大量使用監控相機，但不管在佈建或是運轉的過程中，需要一定程度之人為介入。尤其是佈建初期，需要使用人工進行相機設定校正與道路區域設定，正確之道路區域切割對後續擷取參數與事件有很大的影響，自動化處理有助於大量節省人力成本。本計畫旨在建立一可達到高度自動化並處理異質性道路之智慧型視訊監控系統，並支援動態佈建行動式監控攝影機，自動建立多相機關聯性拓樸。在佈建中達到自動選取感興趣區域，車道數切割，與路口偵測定位，減少所需之人為介入。挑戰性在於，道路與路口走向和形狀以及相機架設之可能性變化極大，增加自動分析的難度。在交通參數擷取方面，利用強韌之特徵擷取與可適性分群演算法，設計在不須重新訓練之前提下，達成不同路段與不同攝影機角度之車流量估測機制。不須針對不同路段進行重新訓練，可以大幅減低繁瑣的收集訓練樣本和標記的工作，增進實際佈建的可行性。對於路口監控，分析號誌時間並評估對其對佇列長度與等待時間之影響，並分析不同路口的交互影響，同時整合多路段異質性道路之參數與事件。另外，在各路段中偵測特定車輛，根據相機所佈建之相對位置做預測與追蹤，達到車輛之再識別與特定車輛之多路段行車路徑重建。 ;Extracting traffic parameters and detecting events with surveillance cameras is an important topic for traffic management in intelligent surveillance systems. Although surveillance cameras are prevalent, human intervention is inevitable during deployment or operation. Camera calibration and settings of road areas in the surveillance scenes are often necessary, especially for early deployment of surveillance cameras. Accurate road area recognition and segmentation are crucial for subsequent traffic parameter extraction and event detection. Automatic processing and setting is helpful for reducing the cost of human resource. This project aims to establish a highly automated intelligent surveillance and traffic management system for integrated heterogeneous roads. The system can deal with surveillance scenes of highway, freeway, or cross-roads. Also, it supports dynamic deployment of mobile surveillance cameras. To eliminate human intervention, the proposed system is designed to achieve automatic region of interest selection, lane segmentation, and road orientation detection. The challenges of the above mentioned tasks include the variety of possible shapes and orientations of the surveillance scenes. For traffic parameter extraction, we propose robust feature extraction and adaptive clustering algorithms to design a flow estimation mechanism that does not require retraining for different camera settings. It substantially reduces the task of collecting training samples and manually labelling the data. Such design improves the feasibility of deployment in practice. Also, for cross-road surveillance, the system analyzes the length of the traffic lights and evaluate their impact on the queue length of each cross road. The correlation of traffic signals of neighborhood cross-roads is also analyzed. Moreover, the parameters and events across heterogeneous road segments are integrated. Specific vehicles are detected and tracked in different road segments. Vehicle re-identification is performed to reconstruct the path of specific vehicles. In addition, the proposed system makes characterized information decimation recommendation according to the traveling condition of each individual vehicle.
