混合車流及高密度下之車輛影像辨識及軌跡追蹤

Abstract

台灣市區道路主要為汽機車組成之混合車流，由於機車體積小且機動性高，機車騎士常會利用汽車間的空隙進行鑽車，當車輛間隙較小、排列較為緊密時，會使車輛辨識受到干擾，同時當車輛於停等紅燈時，容易產生高密度之混合車流，使車輛較難以被偵測到。此外，台灣市區道路路面有許多標線與標字，如停車格、機車待轉區及左轉專用標誌等，當車輛行駛於背景較為複雜之區域時，容易受到干擾，使車輛較難以被辨識出來，同時也容易將路面標線與標字誤判成車輛，使後續軌跡追蹤效果較差。由於台灣市區道路有著背景複雜與高密度混合車流之特性，容易導致車輛未被偵測到，並偵測出大量非車輛物體，使過去車流影像辨識之研究較不適用於台灣市區道路，因此需要建立新的影像辨識技術來擷取市區道路之完整車輛軌跡。 本研究結合卡爾曼濾波與網路流量模型之特性，提出新的物件追蹤方法，以改善混合車流及高密度情況下車輛偵測的辨識率低而導致車輛追蹤績效不佳的問題。先透過卡爾曼濾波進行軌跡追蹤，以彌補車輛未被偵測問題，再透過網路流量模型串聯車輛軌跡，除錯誤的偵測值之影響。比較單獨使用卡爾曼濾波、網路流量模型及混合模型之影像辨識效果，結果顯示卡爾曼濾波會受到車輛未被偵測之影響，導致產生之軌跡較為片段，同時會將錯誤的偵測值視為目標物體，產生出較多錯誤軌跡；網路流量模型則能夠過濾掉大部分錯誤偵測值之影響，但同時亦過濾掉部分正確之車輛軌跡。混合模型則同時保留兩追蹤方法之特性，於不同車流密度的情景下，混合模型都能夠良好的追蹤出車輛，整體結果以混合模型為最佳，顯示適用於擷取市區道路汽機車車流之軌跡。

In Taiwan’s urban road, the mixed traffic is mainly composed of passenger car and motorcycles. Due to the small size and high mobility, motorcyclist can weave in the stream easily. When different types of vehicles mix together on the road, it may reduce the performance of vehicle detection using computer vision. Besides, there are lots of markers on the road surface. When vehicle pass through these areas, the appearance of vehicles will be changed. It will cause the problem of failing to detect true vehicle and generating the wrong detection. Because of these limitations, the existing techniques of vehicle detecting and tracking have poor performance. Thus, the aim of this study is to develop a new algorithm to extract the trajectory of vehicles in urban road. Our approach is combining two tracking algorithm as a hybrid model to improve the performance. First, we use Kalman filter to track vehicle and predict the trajectory for missing detection. Second, we use the network flow model to connect the position of vehicle. We also compare the performance of Kalman filter, network flow model and our approach. The results show that Kalman filter can reduce the missing detection. However, it cannot handle the problem of false detection and also generates many breaking trajectories. Network flow model can almost remove all the false detection, but it also removes some correct trajectory duo to the missing detection. Our approach combines the characteristics of these two tracking algorithms, which can both reduce the missing detection and remove the false detection. The result shows that our approach has better performance of trajectory extraction trajectory in Taiwan’s urban road.