Landscape 3D visual perception simulation and path planning optimization algorithms based on deep learning

Landscape environments present substantial difficulties for autonomous systems because of issues like vegetation occlusion, rolling terrain, varying light levels, and confusing textures that hinder accurate 3D perception and lead path planners to settle upon local optima or to run slowly. To attack this problem, this paper drinks a stride towards proposing the Landscape Perception-Planning Framework (LPPF), an end-to-end lightweight architecture capable of optimizing perception and planning jointly. LPPF includes a MobileNetV3–Swin Transformer architecture integrated to provide robust monocular depth estimation, construction of StyleGAN2-ADA generated synthetic 3D point clouds in multiple weather conditions for the purposes of generalization, and Proximal Policy Optimization (PPO) planner that dynamically adjusts depth confidence into a cost map for error-aware navigation. LPPF is evaluated using 10,000 synthetic LiDAR frames and 500 real LiDAR frames, achieving an overall score of 0.93, an improvement of 19.2% over DPT using the LPPF framework to process under a 50 ms real-time constraint on an embedded platform. By applying channel pruning and INT8 quantization, the model reduces parameters by 85.2% and increases inference by a factor of 3.21 indicating strong accuracy, robustness, and efficiency for intelligent navigation in complex, resource-constrained landscape environments.