In the current work we propose a method to apply generative AI GAN methodology on a hardware-aware NN for lidar data. In order to improve the network semantic segmentation performance, we introduce a GAN network into the segmentation. The adversarial training encourages the model to generalize better by simulating a diverse range of conditions, improving its robustness in real-world applications. This increases the performance with regard to classical real and simulated data based training approaches. Once the NN is trained, we also apply several compression techniques to optimally reduce the latency of the NN for real-time processing of dense point clouds.

• Lidar
• Embedded NN for dense point clouds
• Generative adversarial networks