3D Object Detection using RGB-D images

Object detection is a classic computer vision problem and a challenging one. Through this project we explore object detection of 3D objects using depth images obtained using Kinect. We trained our object classifier by rendering 3D CAD models of regular household objects with different poses and orientations. For each model obtained we extracted VFH and GRSD feature descriptors. Our classifiers included k-D tree model based nearest neighbor search and an SVM. For testing we performed segmentation and clustering on the captured indoor scene to obtain clusters that are similar to the trained data set. Using chi-square metric we were able to classify the objects and detect their pose.This achieved an average accuracy of 67.8% when tested on six different categories.

Report : Download

Presentation: Download