Detection and Tracking for Moving Cameras
The goal of this project is to develop robust algorithms that can automatically detect and track moving objects in moving cameras. Since such algorithms tend to be computationally intensive, we also want to optimize the accuracy while keeping the computational complexity low to process streaming video in real-time. Severe occlusions, poor object representations, and low-quality data are among the major challenges.
Background & Objective: Detecting moving objects automatically is a key component of many visual content analysis tasks. This technology has many potentially promising applications such as news broadcasting, traffic monitoring, safety and disaster relief maintenance and industrial quality control.
Technical Discussion: Most automatic detection and tracking algorithms are designed for stationary camera setups where priori information and heuristics about the scene and the object motion can be easily incorporated. However, for moving cameras, the object motion is a combined with camera motion. This requires estimation and compensation of camera motion to obtain the true object motion. Along this line, we developed a suite of image registration, feature extraction, background generation and frame-difference based motion detection, evidence accumulation, and inter-frame tracking methods. In low-quality video, the moving objects may be small. We improve the detection performance by integrating backward and forward motion history images, and tracking accuracy by employing an adaptive estimation in multiple likelihood state spaces while selectively applying Markovian prediction schemes such as particle filters.
Future Direction: We are working on general purpose and self-adaptive solutions. We consider extending the possible object classes by incorporating appearance models. We explore various ways of improving the speed by taking advantage of the parallel processors to analyze much higher resolution images.
Contact: Fatih Porikli
Technology Areas:
Imaging
Computer Vision
Modification Date: September 17, 2007
