In our previous articles we talk about various available LiDAR technologies. We quantified the scope of work involved with upgrading the US Road infrastructure. We touched upon the size of the data generated from LiDAR surveying. We also addressed concerns about the accuracy of automated systems vs manual annotation.
Deep learning is more important than ever to automate feature extraction from point cloud data. With heavy industries wading in terabytes of data yet to be analyzed, the ability to focus on the basics can appear to be a lost cause. Teaching computers context is the most viable way to address the breadth of the industry’s mapping requirements.
Project ORBIS
For example, in the UK there is a major initiative to document and tag all of the assets in Network Rail’s infrastructure. Project ORBIS is a dynamic institutional drive towards tracking antiquated infrastructure. A common obstacle to cataloging infrastructure is accounting for changes in assets deployed during different decades – some might have been developed over multiple centuries! A wide variance of objects means that any automatic annotation effort has to take into account changes in specifications over time.
Deep Learning
Learning to ride a bike is Deep Learning
A series of neural synapses establishing a pattern to execute a task is called a neural network. Normally a neural network utilizes feedback to introduce opportunities to improve performance. The feedback comes in the form of a scoring algorithm. This is also similar to the mechanism for natural selection.