Don't get me wrong, but how does your problem-adapted version of DBSCAN constitue a form of machine learning?

Clustering is a form of unsupervised machine learning. Many people are surprised to learn this.

http://en.m.wikipedia.org/wiki/Unsupervised_learning

If you have the Lat-long and trust it, why not extract the street using OpenStreetMap?

Great idea! Synthesizing additional signals to obtain more accurate results is one of the core concepts of context platforms. Pulling in extra data from other maps is something I'd seriously consider if the simplicity of the demo use case wasn't the primary concern.