PDF version of this entire document

Haar and Local Binary Patterns

Based on previous work and some further reading, the program might be fine with the following implementation which takes Haar filters or their newer extension into account. Box diagram for the implementation in place can be seen in Figure [*].

Figure: Flow of the foreseen implementation for collision detection/alerting. Resizing if not done, however, unless a very high framerate is sought after (HD footage would be painfully slow).
Image android-process-flow

Based on the paper titled ``Effective Traffic Lights Recognition Method for Real Time Driving Assistance System in the Daytime,'' Kim et al. use similar tools. The paper is quite recent and it too uses Haar filters. Before looking into this approach I also attempted to use template-based search; it is infeasible, however, because cars come in many shapes and sizes, not to mention colour and illumination, context, etc. The example template image (see Figure [*]) enables me to deal with something that in a mockup-type experiment3 would be easy to test, e.g. by putting the phone/tablet in front of a computer screen with the exact same static image on it.

Figure: The template image used as a reference representing a car
Image audi-s5-cabriolet-back

Roy Schestowitz 2012-07-02