Object Recognition and Tracking for Indoor Robots Using an RGB-D Sensor

In this paper, we extend and generalize our previously published approach on RGB-D based fruit recognition to be able to recognize different kinds of objects in front of our mobile system. We therefore first extend our segmentation to use depth filtering and clustering with a watershed algorithm on the depth data to detect the target to be recognized. We forward the processed data to extract RGB-D descriptors that are used to recoup complementary object information for the classification and recognition task. After having detected the object once, we apply a simple tracking method to reduce the object search space and the computational load through frequent detection queries. The proposed method is evaluated using the random forest (RF) classifier. Experimental results highlight the effectiveness as well as real-time suitability of the proposed extensions for our mobile system based on real RGB-D data.