發展以卷積神經網路為基礎之車輛偵測與分類

Vehicle Types Classification and Counting Using Image and Convolution Neural Network Techniques

吳建達(Jian-Da Wu) ； 林育豪(Yu-Hao, Lin)

《車輛工程學刊》 13期 (2019/06) Pp. 21-31

摘要

本研究主要是利用影像處理及卷積神經網路技術來探討道路之交通流量。交通流量的定義主要是參考交通部在2010年所編撰的交通工程手冊，其中提出的小客車當輛數作為道路交通流量的依據。小客車當輛數主要是將車子分類為大型車、小型車、機車／速克達，透過相對小型車之道路容積的比例做計算得到小客車當輛數，也是本研究所探討交通流量的定義部分。本系統主是使用VOC2007所提供的車子種類的圖片資料庫經由卷積神經網路進行預訓練，接著參數調整、資料增廣後獲得所要的訓練模型，然後將由側方拍攝的交通流量的影片隨機截圖後，輸入已經訓練好的模型來進行車輛的車型分類和標記，將車子的邊界使用矩形框框起進行標計並且在矩形框的左上呈現車子的類別，最後使用meam Average Precision（mAP）來評估模型訓練整體的表現。

關鍵字

卷積神經網路 ； 交通流量 ； 車輛分類

並列摘要

Traffic flow is one of the most important information for intelligent traffic transportation engineering. This paper proposed to develop a vehicle classification system for vehicle type using neural network technique. In the early stage of the study, the recognition of different vehicles by using static images was studied. In the secondary stage, the classification and counting system of vehicles by using traffic monitoring are proposed. The study architecture is divided into two parts. Firstly, all types of vehicle picture divided into motorcycles, sedans, recreation vehicles, buses and trucks to build a contrast database. The median filtering and edge detection are used to de-noise action for improving recognition efficiency. The second stage is the data processing in the previous stage as the database of system identification, and then all the data created by the database are input into the classifier for calculation, and finally obtained the recognition rate. The third stage is traffic condition monitoring, and then the number of vehicles is calculated in the lane by using the multi-target tracking method. In this study, the convolution neural network (CNN) of deep learning are used in classification system.

並列關鍵字

convolution neural network ； traffic flow ； vehicle types classification