A deep dive into optimizing a full-stack perception engine using YOLOv8 and SORT, achieving 127 FPS through TensorRT and FP16 quantization.

This post explores road segmentation using a U-Net architecture trained on monocular images from the KITTI dataset. The task focuses on identifying drivable areas from RGB inputs, a challenge due to lighting, shadows, and variable road conditions. I detail the model architecture, training strategies, and key hyperparameters such as learning rate selection and loss functions.