Efficient Object Detection and Tracking Using a Novel MSER-Based Approach
StatisticsView Usage Statistics
This dissertation introduces a novel, real-time, color-based, Maximally Stable Extremal Region (MSER) detection and tracking algorithm. Our algorithm combines MSER-evolution with image-segmentation to produce maximally-stable segmentation. This is achieved using a dual-pass segmentation algorithm that first clusters pixels into a hierarchy of detected regions using an efficient line-constrained evolution process, and then fills gaps between regions to produce dense image segmentation. The resulting region-set offers several unique advantages including fast operation, dense coverage, and temporal stability. It also facilitates efficient computation of additional features including line segments, corners, contours, color histograms, and others. Region tracking is accomplished by matching sub-regions identified around each region's perimeter to those in subsequent frames. Foreground segmentation is achieved by identifying regions that display consistent motion that differs from the background. A static background can improve results but is not required. If models already exist to describe an object, those models can be used to identify foreground regions directly. Regions are modeled and classified using simple color histograms compiled from the contained pixel-clusters. Simple predefined interactions are identified between people and objects using proximity measurements and relative motion vectors. If the interaction information suggests that a theft or other threat might be occurring, the system is capable of dispatching one or more automated robots to investigate or pursue an individual. The system was demonstrated on physical robots that were used in both surveillance and assistance-type scenarios.