臺灣博碩士論文加值系統

長久以來﹐影像處理技術於交通領域的應用均著重在路段車輛的偵測
﹐對於車輛運作複雜的路口﹐確鮮有研究提出﹐但這並不表示路口的重要
性低於路段。相反地﹐路口車流資訊決定號誌之時制計劃(Time
Planning) 亦為道路幾何設計之參考依據﹐對交通的影響極為深遠。
本研究首度將光流理論引入交通領域﹐進行車輛追蹤﹑計數等技術之開發
。由於其相較於傳統影像處理而言﹐對光度較不敏感﹐故偵測結果之可靠
度因此提升。 為使移動場(光流)之求算更精準﹐本研究於前處理先選
擇特徵點─角點進行處理﹐再針對角點以區塊匹配的方式求得速度場。此
過程除可增加光流求算的精確度外﹐並克服孔徑(aperture)問題的發生。
本研究所發展之自動化偵測車輛轉向參數計有軌跡﹐車輛種類﹑轉向交通
量以及車輛瞬間速度等四項輸出。其中車輛種類辨識之驗證結果方面﹐大
型車與機車均無錯誤發生﹐小型車辨識之正確率則為92.65%。轉向交通量
之驗證結果方面﹐直行車輛計數之正確率﹐大型車為80.00%﹐小型車
為92.68%﹐機車則為87.50%；而右轉小型車之計數則無錯誤發生；至於車
輛瞬時速度之誤差則約為 6.86%。不過﹐本研究所測試之數量並不夠多﹐
故模式真正之績效有待未來增加測試數量後﹐方可斷言。

In this thesis, we try to develop an algorithm to detect
turning parameters of vehicles in intersection. We track twelve
directions of vehicular movement and classify three kinds of
vehicle type. Our approach consist of the following main steps:
First, we use Sobel filter to find the edge. Then, we extract
the contour of cars with dilation and erosion procedures. In
order to avoid the aperture problem and acquire accurate optical
flow estimation, the optical flow vectors are computed only on
the image patch centered at corners. After that, we clustered
the optical flow which belongs to the same vehicle by its blob
image. Then we can track its position via velocity field. The
detection system can output four parameters including
trajectory, vehicle type, traffic volume and speed. A case study
was demonstrated through this technique with relatively
satisfactory identification rate of 92.65 in passenger cars.
About flow, accurate rates of go straight trucks is 80.00%,
passenger cars is 92.68%, motorcycles is 87.50%. Accurate rates
of left turn passenger cars is 90.48%, motorcycles is 80.00% and
there is no error occur about right turn passenger cars. The
error of speed is about 6.86%. The performance of the
algorithm,though,is fine,but it needs more cases to calibrate.
After do that, we can finally give judgement on the algorithm.