GMM與EM應用於路測雷達偵測器三車種學習演算法的研發

Abstract

本篇文章主要提出路側雷達偵測器(Road-side radar detector)應用於多車道的環境下所需要的車種分類器 (vehicle type classifier)。本研究以電壓訊號作為輸入變數，並進一步透過傅利葉轉換成頻譜訊號。利用車輛通過偵測區域時，所得到的特徵值，進行學習模型的樣本。 車種分類器的學習架構包含的模型與演算法，使路側雷達偵測器能依據現實的道路環境，偵測通過的車輛來獲得學習樣本，並即時得到通過的車輛是位於哪一個車道的車種資訊。其中，車種主要區成摩托車，小型車，大型車三類。分類所使用的統計模型為二維的高斯混合型，並利用期望最大化演算法(EM algorithm)求解模型參數。

Counting traffic in a single lane is a basic task that can be achieved by using traffic detectors to detect passing vehicles, but it is difficult for road-side radar system to simultaneously detect different vehicle types in multi-lane environments, since the signals reflected from passing vehicles in a single lane influence neighboring lanes. The spread of reflected signals created difficulty in accurately identifying lane boundaries, and leaded that a vehicle classifier in multi-lane situations is in the experimental stage. The aim of this research is to provide a real-time vehicle type classifier in multilane situations based on the lane boundary results. An on-line learning procedure is proposed to form a vehicle classifier. Such a vehicle classifier will utilize the results of on-line automatic lane boundary estimator, and distinguish the vehicles including motorcycles, small-sized and large-sized vehicle. GMM is applied to form a learning model, and integrates an EM algorithm to maximize the likelihood. The real-world data will be gathered to examine the performance of the vehicle classifier.