基於信息共識濾波器之多相機多目標追蹤及行爲分析系統

Abstract

近年來，隨著生活品質的提升，許多娛樂項目也越來越大衆化，例如冰上滑冰或者輪滑等。但是在人員較繁雜的環境中，一些行爲會對自身或者周圍人造成威脅。在這種情況下就需要工作人員去提醒，但是往往場館較大人員密集，工作人員沒辦法在每個區域内都能進行即時的檢測。而在這些大型場館内又存在不同角度和位置的相機，如何使用監控相機所獲取的影像流進行場館内各目標的追蹤和對其行爲進行分析，便是本文所想要討論的部分。 在本論文中，想要解決的主要問題是如何使用多臺相機中的數據，在三維空間中達成一個共識，并且在三維空間中構建姿態估測結果，以在未來能夠精準地對一些動作進行分析。 於是在本論文中，提出了一個使用信息共識濾波器進行信息融合和追蹤的系統。首先使用二維的姿態估測結果進行追蹤以及匹配每個目標的觀測值。然後在三維空間中進行重建以及再一次的信息融合。獲取準確的姿態估測位置。然後進行行爲分析所相關的計算。在該方法中，并不直接使用三維姿態估測的方法，以減少使用三維追蹤和構建估測結果的計算量以及減小對設備算力的要求。 最後，本篇論文展示了實驗結果來討論所提出的方法的有效性。

In recent years, with the improvement of the quality of life, many entertainment projects have become more and more popular, such as skating or roller skating. However, in these cases with a large number of people, some behaviors will pose a threat to themselves or the people around them. In these cases, the staff in venue needs to remind and forbid these actions. However, the scalers of these activities are often big and staffs can’t carry out real-time detection in every area. In these large venues, usually, contain many cameras of different angles and positions to picture the frame in the venues. How to use these images obtained by the surveillance cameras around the venues to track and analyze the behaviors of each target in the venue is the part that this thesis wants to discuss. In this thesis, the main problem to be solved is how to use data from camera network to reach a consensus in 3D space, and construct pose estimation results from 2D to 3D space, so as to accurately perform some actions in the future analysis. So in this thesis, a system using information consensus filters for information fusion and tracking is proposed. The 2D pose estimation method is used firstly. And then using the filters to track and match the observations of each target. Reconstruction and information fusion again are then carried out in 3D space. Get an accurate pose estimation position. The calculations associated with the behavioral analysis are then performed. In this method, the method of 3D pose estimation is not directly used, so as to reduce the calculation amount of using 3D tracking. Moreover, without using 3D pose estimation method also can reduce the requirement on the computing power of the device. Finally, this thesis presents experimental results to discuss the effectiveness of the proposed method