Seeing the trees before the forest [natural object detection]

In this paper, we propose an algorithm that detects and locates natural objects in an outdoor environment using local descriptors. Interest points inside images are detected with a difference of Gaussian (DoG) filter and are then represented using scale invariant local descriptors. Our algorithm learns objects in a weakly supervised manner by clustering similar descriptors together and using those clusters as object classifiers. The intent is to identify stable objects to be used as landmarks for simultaneous localization and mapping (SLAM) of robots. The robot milieu is first identified using a fast environment recognition algorithm and then landmarks are suggested for SLAM that are appropriate for that environment. In our experiments we test our theory on the detection of trees that belong to the plantae pinophyta (pine family). Initial results show that out of 200 test images, our classification yields 85 correct positives, 15 false negatives, 73 correct negatives and 27 false positives.