Background
The use of pan tilt zoom (PTZ) cameras in transportation systems is prevalent because it allows for operators to see a wide range of perspectives. Camera analytics that use PTZ cameras have previously been restricted to a fixed number of preset locations due to the complexity of configuring each new perspective. This research focuses on eliminating the manual process and automating it to allow for an endless number of perspectives which can now be used instead of a fixed number. This software was targeted to work with the Active-Vision™ vehicle analytics software.
Approach
For this research we created a three-step process for automatically calibrating the system each time the camera moves. The first step was to track vehicles, clean the data, and then group the tracks into distinct regions. The second step is to produce a list of candidate orientations which are consistent with the geometry of the camera. The final step evaluates the candidate solutions and chooses the highest scoring one. Steps two and three are then repeated until a desired calibration is produced.
Accomplishments
This research resulted in a novel technique that can reorient a camera once moved and relies on a minimal amount of information to do so. Figure 1 shows an example of a traffic camera on SwRI campus being projected into the top-down perspective needed for analytics. This research has been incorporated in the Active-Vision software and is being used for customer camera feeds.
Figure 1: (left) picture from traffic camera on campus. (right) Traffic camera image projected into GPS coordinates and overlayed onto a satellite top down image.