PDF version of this entire document

Planned Implementation

Originally and principally, my goal is to detect the back of cars. There may be multiple objects in the scenes that need detecting as such, or none. The baseline object from which to calculate distances to nearby rears of cars is marked in green hues.

Rather than taking a lot of movies with a phone and running experiments solely on those, I used a few static images and several videos from YouTube (dashboard-mounted cameras in an urban setting). I trigger for collision warnings typically when the size of the object in front of the car seems too great relative to prior frames (there are variables to keep track of distance changes). Sound gets used if that distance is beneath a particular (predefined) threshold. There are several different sounds based on perceived alert severity.

At present, the program is not features-rich and it lacks accuracy because the classifier was trained on a relatively small-sized annotated set. Training and targets will be explained in the section about experiments, but it is worth noting that the program now latches onto features using local binary patterns rather than the sliding window, template-based approach, which slowed things down a great deal. Framerate is steady at around 5 frames per second, for now, depending on the hardware at hand.