Vision based automatic road traffic density estimation for both stationary and free flowing traffic scenes, and vehicle classification using an ensemble pattern classifier

Abstract

Automatic road traffic density estimation and vehicle classification are very important aspects of today's Intelligent Transportation Systems (ITSs). Traditionally loop sensors have been used for this purpose, but lately vision based systems have been preferred due to their advantages and the problems associated with loop sensors. Many vision based vehicle detection and classification algorithms for free flowing traffic have been proposed. These systems are largely dependent on either motion detection or more generally background modelling and subtraction. There is little reported of traffic scenes with very slowly moving or stationary vehicles for which motion detection based approaches are impractical. This thesis presents a novel vision based road traffic density estimation and vehicle classification approach that is independent of motion detection and background modelling and subtraction. It combines selected image processing, computer vision and pattern recognition algorithms to obtain the traffic parameters. The approach is applied to standstill or slow moving traffic (more generally, 'stop - go' traffic), and free flowing traffic under different illumination conditions during daytime. The proposed system is based on a novel approach in which vehicles are simultaneously extracted from collected video frames and their image negatives through image segmentation using the Laplacian of Gaussian (LoG) edge detector and morphological filtering of the resulting binary image objects. The obtained objects are classified into small, medium and large vehicles on the basis of size using four different classification algorithms and their ensemble. The four base classifiers used are: the k-nearest neighbours (knn), the nearest centroid, the naïve Bayes and the multilayer neural network classifiers. The proposed approach can be used for both stationary and free flowing traffic in contrast to motion detection based approaches. To develop and test the proposed system, video data from a road section was collected using a 5 megapixel camera mounted above the road on which the subject vehicles passed. In order to assess the performance of the system under various illumination levels during the day, three datasets for the free flowing traffic were collected at three different time periods of the day. These datasets were subjected to the vehicle detection process independently. They were finally combined to form an overall dataset for the free flowing traffic which was then subjected to the pattern classification algorithms. In addition to these, a single dataset was collected from a slow moving / stationary traffic scene (‘stop-go’ traffic scene) and used to assess the performance of the system in such traffic scenes. The proposed algorithm was implemented in MATLAB R2015a and an average vehicle detection accuracy of 96.0% and average vehicle classification accuracies of 90.9%, 89.4%, 91.1%, 89.3% and 91.8% by the k-nearest neighbours, the nearest centroid, the naïve Bayes, the multilayer neural network and the ensemble classifiers respectively were achieved for the free flowing traffic dataset. For the slow moving traffic dataset, average detection accuracy of 82.1% and vehicle classification accuracies of 80.4%, 77.2%, 75.9%, 77.7% and 82.2% by the k-nearest neighbours, the nearest centroid, the naïve Bayes, the multilayer neural network and the ensemble classifiers respectively were achieved. The lower percentages for the later dataset were mainly due to occlusions. The main novelty of this thesis is the development of a vision based vehicle detection algorithm that is capable of extracting vehicles from both stationary and fast moving traffic scenes.