Skip to main content

ECE MS Thesis Presentation by Xiao Zhang, via Zoom

Tuesday, May 04, 2021
11:15 am
Floor/Room #: 
via Zoom (please email Xiao Zhang ( for the zoom link if you would like to attend)


Logic Design of Point Cloud Based Convolutional Neural Network Accelerators



LiDAR-based point cloud convolutional neural networks have been widely used in recent 3D object detection tasks. However, for the model inference in edge applications, the performance of a system is restricted from complex model structure, dense computational processing and limited resources of the platform such as power supply and memory space. Therefore, a hardware implementation of the inference model is a proper solution for reaching the high accuracy under a relatively low power consumption and latency. For FPGAs and ASICs architecture, a logic design for the system is a prerequisite procedure. In this thesis, we re-build the inference path of a popular point cloud neural network: VoxelNet. Based on the algorithm we propose a system-level logic design for hardware implementation, including design of Processing Elements (PEs), on-chip input/output buffers, AXI4 based external memory read/write logic and system controller by Finite State Machine (FSM). After verifying the function of Convolution and Fully Connected (FC) layers of logic design. Finally, under behavioural simulation, we implemented and verify 3 × 3 2D convolution and FC layer for an intermediate voxel feature tensor provided by algorithm model in int8 quantization with 4 in-channel and 4 out-channel parallelized processing under the external memory bandwidth of 32-bit.


Research Advisor:

Prof. Ziming Zhang

ECE Department, WPI


Research Committee:

Prof. Xinming Huang

ECE Department, WPI


Prof. Kaveh Pahlavan

ECE Department, WPI