- Automated feature classification/extraction is not an easy problem to solve.
- A pixel by pixel approach is not the way to go.
- A multi-sensor, data fusion approach seems to hold more promise.
- This is actually a remote sensing, image processing problem
As noted in my last blog the goal is not be in the laser scanning business, it is to use this disruptive technology to solve real world problems – more efficiently and hopefully with superior 3D results. In this regard, one of the most important challenges facing our industry is automating the classification and extraction of features, or more specifically 3D objects from the point cloud. As more than one industry expert has told me, this is akin to the “Search for the Holy Grail”.
Most of the approaches to solving this image processing problem, not only in the laser scanning world, but in the larger universe of remote sensing, tackle this problem on a pixel by pixel basis. One company that I used to work for created an ERDAS Imagine plug-in that was called the “Sub-Pixel Classifier”. Their theory was that all pixels are “mixed” pixels, and they attempted to determine the materials that made up the pixel. Since they would not reveal their methodology, the product had limited success.
In a previous blog from my recent trip to the ILMF conference I noted the unique approach being used by Definiens. Instead of interrogating the scene pixel by pixel, they use an iterative approach to develop an understanding of the pixel in context with its surrounding pixels. A key part of their strategy is to make use of data from other sensors in their analysis. This can include orthophotography, including infrared bands, multi and hyperspectral. I was told they can use any other available image information.
I believe in this multi-sensor, data fusion approach to automating the feature classification and extraction process. I think it is critical to achieving the kind of results that customers have a right to expect, and that it would be of great benefit to advancing the use of laser scanning. A colleague at a large aerial LIDAR mapping firm tells me that the state of the art for extracting planimetrics on a LIDAR project is still heads up digitizing.
This is fertile ground for software development, and I don’t think we should look to the equipment vendors to create these solutions. At the core, this is a remote sensing/image processing problem, not a laser scanning problem. As customers we need to create the demand for these next generation solutions, hopefully with a user interface that the average technician can understand.