利用模糊類神經網路加強彩色影像車輛偵測

Abstract

在交通控制與規劃中，交通參數是規劃者進行決策分析時重要的參考依據。近年來，車輛影像辨識逐漸取代傳統線圈偵測以擷取交通車流參數。在進行影像偵測與辨識時，「背景相減法」是普遍被研究者所使用的，但過去多是應用在灰階影像的辨識上，缺乏在彩色影像方面的研究。但由於彩色影像辨識具有R、G、B三種像素，因此能夠比灰階影像提供更多的資訊作為分析基礎原件。 為提升影像辨識的準確率，本研究結合模糊類神經網路方法與背景相減法，並採用橫跨單一車道之線偵測器，發展彩色影像車輛偵測模式。在模式構建方面，本研究設計一四層之模糊類神經網路架構，並以倒傳遞演算法進行網路的訓練。樣本蒐集係針對市區道路與高速公路，於不同光線情境下進行影像拍攝。本研究在虛擬偵測線上分別放置三點、五點與七點偵測器，藉此比較不同偵測點數對於偵測效果的影響。除此之外，本研究亦將所發展之彩色影像偵測器與傳統之彩色影像偵測器所得之效果進行比較。實證分析發現，以七點偵測點為例，結合模糊類神經網路的彩色影像車輛偵測系統，無論在市區道路或高速公路，白晝或夜晚，所得的車輛偵測準確率皆約達90%以上。五點與七點虛擬偵測器所得之偵測效果較三點偵測器為佳，五點與七點虛擬偵測器所得之偵測效果則無明顯差異。結合模糊類神經網路的模式所得的偵測效果比傳統的車輛偵測模式為佳。探究本系統偵測失敗的主因，在日間主要是因為車輛顏色與道路背景相似之故。在夜間方面，主要是因為受到車頭燈以及車輛變換車道的影響。

Traffic data collection is extensively required by planners in traffic control and management. In recent years, traffic parameters are obtained by vehicular image detection system instead of loop detectors. While utilizing for vehicular detection, background-differencing technique is usually applied by researchers. The image of pervious researches is based on gray-level, not color-level. Owing to the color image that consists of R, G, B pixels, color image provides more information than gray-level image. In order to enhance the detection accuracy rate, this paper develops a color image vehicular detection (CIVD) system by incorporating background differencing method with fuzzy neural network (FNN). A pseudo line detector is placed on the monitor to detect a vehicular image. A four-layer neural network is constructed and the network parameters are trained by backpropagation algorithm. Traffic flow scenes under different road environments with various lighting conditions are tested. Three-, five- and seven-points pseudo detectors are installed on the screen to compare the detection performance. Besides, performances of this new detection system and a traditional detection system are compared. It shows that the detection accuracy rates for this FNN CIVD system with seven pseudo detection points in urban streets or freeways, daytime or nighttime, can reach 90% or over. The performances of five- and seven-points pseudo detectors are better than that of three detectors. The differences of detection accuracy between five and seven-points detector are slight. The performance of this detection system incorporating with fuzzy neural network is better than that of the traditional detection system. Main reasons of failure detection are discussed. In the daytime, it is due to a resemblance between most gray vehicles and the roads. In nighttime, it is affected by the headlights and lane-changing vehicles.