With the the ubiquitous deployment of surveillance cameras, huge
amounts of video data are being generated at every moment. Analyzing
the massive surveillance videos in an efficient manner has become a
pressing task. Visual object tracking is one of the enabling
technologies for video analysis and has received much attention in the
computer vision community during the last decade. Despite the recent
advances in the visual tracking research, there are still several
challenges to the existing methods such as efficiency, accuracy,
resilience to visual ambiguities, etc. To address such challenges and
improve the tracking performance, the constraints specific to the
surveillance applications need to be utilized, which have not been
thoroughly studied before. The objective of this dissertation is to
exploit the constraints pertaining to the surveillance applications
and integrate them into the probabilistic tracking framework for
effective visual tracking.
This dissertation first presents the integration of environment
constraints into the particle filtering framework for effectively
tracking objects for the urban surveillance applications. In these
applications, the movements of objects are constrained by structured
environments. Therefore, the relationship between objects and
environments can be exploited as additional information for improving
the performance of tracking. An environment state is introduced to
represent the relationship between the objects and
environments. Distance transform is used to model the environment
state. The adaptive dynamics model and environment prior are devised
for the particle filter to fully utilize the environment information
in the tracking process.
Then the integration of electronic localization for effective visual
tracking is studied. Electronic signals, like cellular, WiFi and
Bluetooth signals emitted from mobile phones, are ubiquitously present
and can be associated with the objects of interest. A directional
antenna is used for collecting the signals and performing rough
electronic localization. Such location information is fed into the
visual tracking algorithm as object motion constraints, so the
uncertainty and search space of visual tracking are significantly
reduced.
Finally, a stereo tracking method for measuring the speed of a moving
vehicle within a structured environment is presented. The stereo
constraint between the two views and the path constraint for the
vehicle's motion are exploited for accurate visual tracking which
overcomes the limitation of depth accuracy in long range stereo. In
the proposed method, visual stereo tracking and motion estimation in
3D are integrated within the framework of particle filtering. The
visual tracking processes in the two views are coupled with each other
since they are dependent upon the same 3D motion and correlated in the
observations. Considering that the vehicle's motion is physically
constrained by the environment, the path constraint reconstructed from
stereo views is utilized to reduce the uncertainty about the vehicle's
motion and improve the accuracy for both tracking and speed measurement.