Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Video analysis is an active and rapidly expanding research area in computer vision and artificial intelligence due to its broad applications in modern society. Many methods have been proposed to analyze the videos, but many challenging factors remain untackled. In this dissertation, four statistical modeling methods are proposed to address some challenging traffic video analysis problems under adverse illumination and weather conditions.
First, a new foreground detection method is presented to detect the foreground objects in videos. A novel Global Foreground Modeling (GFM) method, which estimates a global probability density function for the foreground and applies the Bayes decision rule for model selection, is proposed to model the foreground globally. A Local Background Modeling (LBM) method is applied by choosing the most significant Gaussian density in the Gaussian mixture model to model the background locally for each pixel. In addition, to mitigate the correlation effects of the Red, Green, and Blue (RGB) color space on the independence assumption among the color component images, some other color spaces are investigated for feature extraction. To further enhance the discriminatory power of the input feature vector, the horizontal and vertical Haar wavelet features and the temporal information are integrated into the color features to define a new 12-dimensional feature vector space. Finally, the Bayes classifier is applied for the classification of the foreground and the background pixels.
Second, a novel moving cast shadow detection method is presented to detect and remove the cast shadows from the foreground. Specifically, a set of new chromatic criteria is presented to detect the candidate shadow pixels in the Hue, Saturation, and Value (HSV) color space. A new shadow region detection method is then proposed to cluster the candidate shadow pixels into shadow regions. A statistical shadow model, which uses a single Gaussian distribution to model the shadow class, is presented to classify shadow pixels. Additionally, an aggregated shadow detection strategy is presented to integrate the shadow detection results and remove the shadows from the foreground.
Third, a novel statistical modeling method is presented to solve the automated road recognition problem for the Region of Interest (RoI) detection in traffic video analysis. A temporal feature guided statistical modeling method is proposed for road modeling. Additionally, a model pruning strategy is applied to estimate the road model. Then, a new road region detection method is presented to detect the road regions in the video. The method applies discriminant functions to classify each pixel in the estimated background image into a road class or a non-road class, respectively. The proposed method provides an intra-cognitive communication mode between the RoI selection and video analysis systems.
Fourth, a novel anomalous driving detection method in videos, which can detect unsafe anomalous driving behaviors is introduced. A new Multiple Object Tracking (MOT) method is proposed to extract the velocities and trajectories of moving foreground objects in video. The new MOT method is a motion-based tracking method, which integrates the temporal and spatial features. Then, a novel Gaussian Local Velocity (GLV) modeling method is presented to model the normal moving behavior in traffic videos. The GLV model is built for every location in the video frame, and updated online. Finally, a discriminant function is proposed to detect anomalous driving behaviors.
To assess the feasibility of the proposed statistical modeling methods, several popular public video datasets, as well as the real traffic videos from the New Jersey Department of Transportation (NJDOT) are applied. The experimental results show the effectiveness and feasibility of the proposed methods.
Shi, Hang, "Novel statistical modeling methods for traffic video analysis" (2021). Dissertations. 1541.