Deep Convolutional Neural Network Design Approach for 3D Object Detection for Robotic Grasping

Automation is increasing with the advent of 3D- and depth images using an infrared camera. Formerly, most of the object detection and recognition were done by a 2D-camera of Red-Green-Blue (RGB) images. Today, with the availability of economical 3D-sensors, people started investing in 3D-object detection. Over recent years, Convolutional Neural Network (CNN) has reached the epitome of image classification for different application. Explicitly in 2D-CNN, there is massive progress for object detection, but for 3D-CNN, it is still at the beginning of an era. In this proposal, RGB images and depth images will be used to classify the objects using a Structure Sensor camera. Previously, researchers mainly worked on 2D-images and received considerable accuracy, but the accuracy decreases for 3D-image classifications using 2D-techniques. The complexity of the proposed model is increased due to the partial images with variation in illumination, occlusion, and cluttered images. To overcome the complexity problem, Deep Convolutional Neural Network (DCNN) is proposed with the usage of RGB-D images, which includes one 2D-image in RGB, and the other in-depth image form. Two methods are proposed, one will be using two parallel DCNN model and another method will be using three parallel DCNN model. Afterwards, the parallel models need to be concatenated to get 3D-object detection. Public-available 3D-dataset will be used to evaluate the model, and the results of both of the models will be compared.