Sample Identification (WIP)

The purpose of this activity is to determine the distance limits on seeing and identifying a sample. Each type of known sample is used. 

The original photos are taken with a Panasonic Luminx digital camera at full resolution. Infranview is then used to reduce the image to a 640x480 pixel image for processing using RoboRealm. Other research will determine if this size image is adequate for the search or whether a larger image is needed. Or, it may be that a smaller image may be usable. There will be a trade off among frame rate, i.e. how often a new image is available, processing, and the image size.

A set of images is collected at increasing distances to see how far the 'bot can be and still reliably detect the sample. The distance are approximately 1, 2, 4, 6, 8 and 10 meters. (I just paced off 6 feet between each shot and added a little more to approximate a 2 meter distance. This is research for analysis to see what might be possible so precision is not required.) The camera is on a mono-pod at a height of 1.5 meters.

Orange Pipe

Orange pipe sample on gravel and grassThe orange pipe samples is a 20 cm long 3/4 inch PVC pipe painted fluorescent orange. Three of these were made and laid near one another. As you can see in the image one is on gravel, one on grass perpendicular to the line of sight, and the last one at an angle to the line of sight. This image is from 2 meters. 

The image was taken on a day with intermittent rain at a time when it was not raining and clouds obscured the sun. This was intended to be a worst case scenario since, presumably, a sunny day would make the sample, especially an orange one, easier to see. However, more intense sunlight may increase the false positives as the intensity is stronger on other areas of the image. All this will be tested later. 

A simple test using an RGB filter and a Blob Size Filter was used. At 8 meters the processing is able to see two little orange specs which would not be sufficient to detect the samples. Closer in at 6 meters the pipe at an angle is not detected. At 4 meters and under the pipes are detected. As the distance increases the number of false positives increases. 

The Area Analysis found that a 3 to 6 meter search area for a camera would be needed to complete the search in the time required. The results for the orange pipe indicate this may be possible although if the pipe is not positioned ideally it may be more difficult. Since the vision processing used in this test was pretty simple it is likely more rigorous vision processing would see the sample.