Overview

Scalable, cost-effective generation and updating of high-definition maps (HD-maps) is a foundational challenge for the widespread deployment of autonomous vehicles. The quality of any HD-map depends on the accuracy of the underlying sensor calibration and localization pipelines. My Ph.D. research tackles this end to end: from rigorous camera modeling, through multi-sensor calibration, to geometrically consistent large-scale LiDAR–inertial mapping. It culminates in a low-cost prototype mobile mapping system that can be driven or worn as a backpack.

• Institution: University of Calgary, Schulich School of Engineering, Department of Geomatics Engineering
• Degree: Doctor of Philosophy, Geomatics Engineering
• Year: 2026

Thread 1: Beyond Checkerboards — Photogrammetric Camera Calibration

Checkerboard-based calibration is popular because it is simple and easy to deploy, and it is used even in mission-critical computer vision pipelines. But it can't model high-precision camera geometry well, which hurts downstream tasks that need geometric accuracy. This work revisits intrinsic and extrinsic camera calibration within the framework of photogrammetric self-calibration (PSC).

• PSC consistently outperforms checkerboard calibration in both accuracy and precision
• Evaluated across 3D reconstruction with structure from motion (SfM), visual SLAM, and novel-view synthesis and reconstruction
• Argues for calibration methods that reflect the physical and projective properties of real camera systems

Publication: Bharadwaj, A. & Lichti, D. D. (2026). Beyond Checkerboards: Advantages of Photogrammetric Camera Calibration for Robust 3D Vision and Novel-View Generation. Remote Sensing, 18(13), 2220. [DOI] [Project Page]

Thread 2: Learning-Based LiDAR–Camera Calibration

Reliable perception and mapping depend on accurately fusing camera imagery with onboard LiDAR and IMU. Camera–inertial calibration is well studied, but LiDAR–camera extrinsic calibration is still hard outside controlled environments. I proposed a two-step, learning-based calibration framework:

1. Coarse estimation: instance-wise correspondences from everyday objects in the scene give an initial estimate of the extrinsic parameters, with no calibration target needed
2. Fine refinement: an end-to-end neural network refines the coarse estimate into accurate calibration parameters

This lets the system keep adapting to changes in sensor pose during real-world operation.

Thread 3: Neural Implicit LiDAR–Inertial Odometry

Neural implicit methods have emerged as an effective alternative to classical LiDAR mapping. They match classical accuracy while needing far less storage for the map. However, their accuracy degrades in highly dynamic scenes, such as areas with pedestrian motion. I developed a neural implicit inertial odometry framework that jointly integrates LiDAR and inertial measurements to produce geometrically consistent 3D point clouds.

• Adaptive LiDAR downsampling to preserve fine geometric detail
• Neural point correspondences for accurate inter-scan odometry estimation
• Attention-based loop closure to keep large-scale maps globally consistent

Prototype Mobile Mapping System and Evaluation

As part of this work, I designed and built a cost-effective prototype mobile mapping system that operates in both driving and backpack configurations, and used it to collect a new dataset. The tools and algorithms were evaluated on public benchmarks (Oxford-Spires, KITTI, and EuRoC-MAV) and further validated on this newly collected dataset.

Publications

• Bharadwaj, A. S. (2026). Cost-Effective Mapping Platform to Generate HD-Maps for Autonomous Vehicles and Digital Twins. Ph.D. Dissertation, University of Calgary. [DOI]
• Bharadwaj, A. & Lichti, D. D. (2026). Beyond Checkerboards: Advantages of Photogrammetric Camera Calibration for Robust 3D Vision and Novel-View Generation. Remote Sensing, 18(13), 2220. [DOI]