Adaptive LiDAR Mapping

Detail where it matters

FOVEA-NET proposes intelligent, variable-resolution 2.5D LiDAR mapping for dynamic environment perception. Inspired by human foveal vision, it allocates more detail to critical risks and the predicted vehicle path rather than relying on fixed distance rings.

A 2.5D elevation map retains height information within a two-dimensional grid. The aim is to preserve useful terrain and obstacle information while reducing the computational and memory demands of uniformly dense mapping.

My contribution and the prototype

I led Team Gol D and personally developed the animated prototype demonstrating adaptive-resolution mapping, terrain analysis, and static and dynamic obstacle perception. The supplied screenshots show the FOVEA-NET driving visualisation, including route guidance, light-beam and grid views, and different camera perspectives.

This is a visual demonstration using simulated geometry, not a physical LiDAR measurement system or validation of the complete proposed AI pipeline. The proposal identifies JavaScript and Three.js for its interactive visualisation.

The proposed perception pipeline

The report proposes acquiring and filtering LiDAR point clouds with Open3D and NumPy, followed by PointNet++ semantic segmentation in PyTorch. Terrain and obstacle classification would use PointNet++ and sparse convolutional networks.

A risk and attention stage would consider the predicted vehicle path, object risk, and uncertainty. Its output would guide a variable-resolution 2.5D elevation grid, using Morton (Z-order) coding for spatial organisation and LiDAR object tracking for temporal memory.

The planned workflow progresses through data ingestion and setup, perception AI, core mapping, and interactive visualisation. These components describe the proposed architecture, rather than a claim that every component is implemented in the animated prototype.

Safety, feasibility and fallback paths

The proposal builds on established tools, pretrained models, and a modular architecture. Adaptive resolution is intended to manage computational load; camera–LiDAR fusion is proposed to address sparse long-range observations.

Morton-coded sparse grids are intended to preserve spatial and height information across resolution boundaries. To mitigate AI uncertainty, the report specifies a fixed 10-metre safety zone with a minimum nearby resolution. This is a proposal requirement, not a measured safety guarantee.

If the deep-learning pipeline is not ready, the proposed fallback combines RANSAC ground-plane extraction with DBSCAN object clustering.

Planned evaluation and research foundations

CARLA is proposed for simulated LiDAR testing, with SemanticKITTI and nuScenes-LiDARseg identified as training and evaluation datasets. The report does not provide measured accuracy, latency, memory savings, or comparative benchmark results.

The cited foundations include PointNet++ (Qi and colleagues), Robot-Centric Elevation Mapping with Uncertainty Estimates (Fankhauser and Hutter), SemanticKITTI (Behley and colleagues), and nuScenes (Caesar and colleagues). Full references are included in the attached report.

Intended impact and future scope

The intended benefits are more focused perception of pedestrians, cyclists, potholes, obstacles, and terrain, with lower computation and memory requirements. These are expected benefits of the proposed system, not experimentally established outcomes.

Future directions include real LiDAR integration, vehicle-to-everything communication, and fleet integration.

The challenge and current status

Prepared for Smart India Hackathon 2026, problem statement SIH26053: Adaptive Variable Resolution 2.5D Lidar Mapping for Dynamic Environment Perception, under the Smart Vehicles theme and Software category. The proposal identifies Team Gol D, team ID 143452.

Our team won the college-level selection and submitted the project to the next stage. Next-stage results are awaited. The original Gol D proposal is available through Read technical report, and the prototype video is linked above.