Publicação
Real-time implementation of 3D LiDAR point cloud semantic segmentation in an FPGA
| Resumo: | In the last few years, the automotive industry has relied heavily on deep learning applications for perception solutions. With data-heavy sensors, such as LiDAR, becoming a standard, the task of developing low-power and real-time applications has become increasingly more challenging. To obtain the maximum computational efficiency, no longer can one focus solely on the software aspect of such applications, while disregarding the underlying hardware. In this thesis, a hardware-software co-design approach is used to implement an inference application leveraging the SqueezeSegV3, a LiDAR-based convolutional neural network, on the Versal ACAP VCK190 FPGA. Automotive requirements carefully drive the development of the proposed solution, with real-time performance and low power consumption being the target metrics. A first experiment validates the suitability of Xilinx’s Vitis-AI tool for the deployment of deep convolutional neural networks on FPGAs. Both the ResNet-18 and SqueezeNet neural networks are deployed to the Zynq UltraScale+ MPSoC ZCU104 and Versal ACAP VCK190 FPGAs. The results show that both networks achieve far more than the real-time requirements while consuming low power. Compared to an NVIDIA RTX 3090 GPU, the performance per watt during both network’s inference is 12x and 47.8x higher and 15.1x and 26.6x higher respectively for the Zynq UltraScale+ MPSoC ZCU104 and the Versal ACAP VCK190 FPGA. These results are obtained with no drop in accuracy in the quantization step. A second experiment builds upon the results of the first by deploying a real-time application containing the SqueezeSegV3 model using the Semantic-KITTI dataset. A framerate of 11 Hz is achieved with a peak power consumption of 78 Watts. The quantization step results in a minimal accuracy and IoU degradation of 0.7 and 1.5 points respectively. A smaller version of the same model is also deployed achieving a framerate of 19 Hz and a peak power consumption of 76 Watts. The application performs semantic segmentation over all the point cloud with a field of view of 360°. |
|---|---|
| Autores principais: | Delgado, Pedro Paulo Fontes |
| Assunto: | LiDAR Deep learning FPGA Engenharia e Tecnologia |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | In the last few years, the automotive industry has relied heavily on deep learning applications for perception solutions. With data-heavy sensors, such as LiDAR, becoming a standard, the task of developing low-power and real-time applications has become increasingly more challenging. To obtain the maximum computational efficiency, no longer can one focus solely on the software aspect of such applications, while disregarding the underlying hardware. In this thesis, a hardware-software co-design approach is used to implement an inference application leveraging the SqueezeSegV3, a LiDAR-based convolutional neural network, on the Versal ACAP VCK190 FPGA. Automotive requirements carefully drive the development of the proposed solution, with real-time performance and low power consumption being the target metrics. A first experiment validates the suitability of Xilinx’s Vitis-AI tool for the deployment of deep convolutional neural networks on FPGAs. Both the ResNet-18 and SqueezeNet neural networks are deployed to the Zynq UltraScale+ MPSoC ZCU104 and Versal ACAP VCK190 FPGAs. The results show that both networks achieve far more than the real-time requirements while consuming low power. Compared to an NVIDIA RTX 3090 GPU, the performance per watt during both network’s inference is 12x and 47.8x higher and 15.1x and 26.6x higher respectively for the Zynq UltraScale+ MPSoC ZCU104 and the Versal ACAP VCK190 FPGA. These results are obtained with no drop in accuracy in the quantization step. A second experiment builds upon the results of the first by deploying a real-time application containing the SqueezeSegV3 model using the Semantic-KITTI dataset. A framerate of 11 Hz is achieved with a peak power consumption of 78 Watts. The quantization step results in a minimal accuracy and IoU degradation of 0.7 and 1.5 points respectively. A smaller version of the same model is also deployed achieving a framerate of 19 Hz and a peak power consumption of 76 Watts. The application performs semantic segmentation over all the point cloud with a field of view of 360°. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.