The work presented in this dissertation provides a framework for object detection,tracking and vehicle classification in urban environment. The final aim is to produce a system for traffic flow statistics analysis. Based on level set methods and a multi-phase colour model, a general variational formulation which combines Minkowski-form distance L2 and L3 of each channel and their homogenous regions in the index is defined. The active segmentation method successfully finds whole object boundaries which include different known colours, even in very complex background situations, rather than splitting an object into several regions with different colours. For video data supplied by a nominally stationary camera, an adaptive Gaussian mixture model (GMM), with a multi-dimensional Gaussian kernel spatio-temporal smoothing transform, has been used for modeling the distribution of colour image data. The algorithm improves the segmentation performance in adverse imaging conditions. A self-adaptive Gaussian mixture model, with an online dynamical learning rate and global illumination changing factor, is proposed to address the problem of sudden change in illumination. The effectiveness of a state-of-the-art classification algorithm to categorise road vehicles for an urban traffic monitoring system using a set of measurement-based feature (BMF) and a multi-shape descriptor is investigated. Manual vehicle segmentation was used to acquire a large database of labeled vehicles form a set of MBF in combination with pyramid histogram of orientation gradient (PHOG) and edge-based PHOG features. These are used to classify the objects into four main vehicle categories: car, van (van, minivan, minibus and limousine), bus (single and double decked) and motorcycle (motorcycle and bicycle). Then, an automatic system for vehicle detection, tracking and classification from roadside CCTV is presented. The system counts vehicles and separates them into the four categories mentioned above. The GMM and shadow removal method have been used to deal with sudden illumination changes and camera vibration. A Kalman filter tracks a vehicle to enable classification by majority voting over several consecutive frames, and a level set method has been used to refine the foreground blob. Finally, a framework for confidence based active learning for vehicle classification in an urban traffic environment is presented. Only a small number of low confidence samples need to be identified and annotated according to their confidence. Compared to passive learning, the number of annotated samples needed for the training dataset can be reduced significantly, yielding a high accuracy classifier with low computational complexity and high efficiency.

The Mobility Analysis Section of the CDOT Division of Transportation Development (DTD) developed this study to determine whether the cluster count method developed by CDOT is statistically reliable for estimating vehicle classification on urban roadways with average daily traffic volumes exceeding 15,000 vehicles per day. Specifically, CDOT needed to assess whether or not the percentages of vehicles in the 13 FHWA vehicle classifications estimated by the cluster count method differ significantly from expected percentages obtained by 24-hour counts. Since vehicle classification is expensive to perform by manual observation over long periods of time, a statistically reliable method of estimating vehicle type percentages on urban roadways using a less time-consuming method is desirable. The study team utilized the chi-square statistical test to evaluate the similarity between vehicle classifications collected using the cluster count method and 24-hour vehicle counts collected using other data collection methods. Vehicle classification data were collected at 12 sites around Denver, Colorado that represented different roadway classes. The statistical tests between the data collected using the cluster count method and the 24-hour counts revealed that the current cluster count method varied beyond an acceptable statistical similarity to the 24-hour counts. Upon reaching this conclusion, the study panel simulated various changes to the short duration count methodology in an effort to identify the greatest improvement in statistical accuracy. As a result of this study, the recommended short duration vehicle classification methodology requires vehicle counts to be performed for 15 minutes every hour for a 24-hour period. This method exhibits strong statistical similarity to the 24-hour classification counts for all roadway classes and study sites included in this analysis. This collection method is statistically accurate, easy for field personnel to understand and collect, and is about onethird of the cost of a manual 24-hour count. The Mobility Analysis Section of DTD has developed a guidebook on the recommended short duration count methodology that will be available to CDOT staff, data collectors, consultants, and other public agencies. This guidebook outlines how to collect the short duration classification data, process and manage the data, and perform quality control checks.