Object classification from segmented LiDAR input using 3 dimensional convolutional neural networks

• Main Author: Pangottil Shanoop
• Other Authors: Justin Dauwels
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/73132
• Institution: Nanyang Technological University

As autonomous vehicles are poised to enter the mainstream in the automobile industry, an important requirement for these platforms is the ability to robustly recognize and react to objects in the real world. This is further compounded by the fact that other autonomous platforms like delivery robots and industrial collaborative systems would have to actively make decisions based on the visual feedback from their sensors. Range sensors such as LiDAR and RGBD are commonly found sensors in modern robotic platforms, providing a richer dataset than any other single sensor platform. Most of the current algorithms for classification and segmentation do not however use the depth data from the 3D data or employ work arounds, often sacrificing classification performance. This thesis is a study into the classification capabilities of 3D convolutional neural networks and evaluates the performance on a 3D CNN implementation [1] in a publicly available dataset [3] and compares it to the state of the art performance metrics as put forward by [2]. This thesis also attempts to find the optimal grid for a voxelization problem by comparing three approaches as mentioned by [1] and verifies the results put forward by the authors. To study these, a 7-layer 3D convolutional neural network based on [1] is used. Slight modifications of the hyper-parameters to accommodate the new dataset is also discussed in this thesis. Finally, the limitations of 3D CNN networks is discussed and its effect on the results of this thesis and improvements as suggested by [15] are also discussed.