Zone Maps for Vehicle Detection

Zheng Zhong, Gierad Laput, and Yi Lu Murphey (2008)

As an undergraduate research assistant for the Intelligent Systems research group at the University of Michigan - Dearborn, I developed an idea to reduce false alarm occurrences that were inherent with the group's vehicle detection system. I observed that some false alarms were detected in sections of the camera viewport where a vehicle is highly unlikely to be spotted (e.g. at the corners).

The system processes image feeds from a mounted camera.
Once processed, the system can detect vehicles from the camera feed.
However, the system makes mistakes and generates false alarms.

I collaborated with my team to gather available datasets and created a "heat" map of where vehicles were likely to be spotted. The heat map was then dissected into "zones." Whenever the system detects a "candidate" vehicle, it will determine which zone the candidate will belong. It will then compare the properties of the zone (which also considers properties such as size and distance from the vanishing point) to decide whether the candidate is a vehicle or not.

Post-processing can be applied to determine which objects are actually vehicles by exploiting their position and size.
A dissected area of the image represents a zone, which stores information such as frequency of occurrence and vehicle size.
A candidate object is then evaluated based on the zone properties to determine whether it is a vehicle.

Besides learning tons of cool new stuff, the opportunity to interact and peer into the minds of the talented individuals in our team made this a wildly exciting experience.

ยป Related Links and Keywords