Perception,
SLAM &
State Estimation

A targetless temporal calibration pipeline that estimates the real-world time offset between an unsynchronized stereo camera and LiDAR using cross-modal edge alignment. Unlike target-based approaches, this method operates directly on natural scene edges without requiring a calibration target in the scene. The pipeline projects LiDAR point clouds onto calibrated camera frames, computes Canny edges on both modalities, and scores alignment using distance transforms — then sweeps over candidate offsets to find the optimum. The estimated 70ms offset was independently validated using Powell optimization and dense grid search cross-validation. IMU preintegration merges 3 consecutive scans to produce denser point clouds, improving calibration robustness under motion.

A real-time 2D vehicle localization system that fuses a 200 Hz IMU with GNSS measurements using an Extended Kalman Filter. The prediction step propagates state using IMU kinematics; the update step incorporates GPS position measurements weighted by their uncertainty. Noise characteristics for both sensors were carefully modeled to tune the process and observation covariance matrices. In GPS-denied zones (tunnels), the filter switches to IMU-only dead reckoning with adaptive covariance inflation, maintaining trajectory continuity until GPS signal is restored.

An online Visual SLAM pipeline deployed on real hardware — a ZED Mini stereo camera mounted on a mobile platform traversing underground GPS-denied tunnels. RTAB-Map running in ROS2 performs real-time loop closure detection, pose graph optimization, and 3D occupancy mapping. The multi-sensor platform also integrates LiDAR and a VectorNav IMU, with careful sensor bring-up, mounting calibration, and coordinate frame transforms across the full sensor suite. Mapping accuracy was validated by comparing reconstructed 3D maps against known ground truth tunnel layouts.

A full Structure from Motion pipeline built from scratch — no high-level OpenCV reconstruction utilities. Starting from 24 monocular images, the pipeline extracts SIFT features, matches them with RANSAC-based outlier rejection, recovers relative camera poses via essential matrix decomposition, and triangulates 3D points. Non-linear bundle adjustment over 1,477 landmarks via GTSAM factor graph optimization reduces reprojection error and improves reconstruction accuracy by 26%. Camera localization in GPS-denied environments is enabled via PnP pose estimation on previously triangulated points.

A from-scratch implementation of 3D Gaussian Splatting applied to a Buddha statue dataset. The scene is initialized from a sparse COLMAP point cloud; each Gaussian is parameterized by position, covariance, opacity, and spherical harmonic color coefficients, all optimized end-to-end via differentiable EWA rasterization. Densification applies clone/split/prune operations every 100 iterations to adapt Gaussian density to scene complexity. A key insight from this project: on a 26-image dataset, the 7,500-iteration checkpoint achieves 31.42 dB PSNR — over 16 dB better than 30,000 iterations, where Gaussians overfit to training views and collapse into needle-like artifacts. The entire pipeline is containerized with Docker and ships with a GitHub Actions CI/CD workflow.

A 2D image mosaicing pipeline for 28 underwater images using GTSAM factor graph optimization. Pairwise homographies are estimated from SIFT feature matches; loop closure constraints on overlapping image pairs are added as factors to eliminate accumulated drift. The factor graph is optimized to produce a globally consistent, seamless mosaic. Robust homography estimation handles the challenging low-light, feature-sparse conditions typical of underwater imagery.

Input Data & Results

An illumination correction pipeline that conditions color constancy on semantic object identity. Rather than applying a global white balance, the system uses YOLOv8 detections to identify each object in the scene and applies per-category chromaticity priors learned from 5,600 COCO-derived images. This object-aware approach produces significantly more accurate color correction under challenging or mixed lighting — improving color accuracy by 63% over the baseline. The pipeline runs at 130ms per image on CPU, covering 80 COCO object categories.

A U-Net based pipeline for joint low-light denoising and exposure correction, trained and evaluated on the LOLv1 benchmark dataset. The architecture was redesigned mid-project to incorporate skip connections that better preserve high-frequency structural detail across encoder-decoder stages. The loss function combines L1 reconstruction loss with a perceptual (SSIM-based) term to eliminate color cast artifacts while maintaining texture fidelity. The final model achieves 19.37 dB PSNR — a 17% improvement over the 16.57 dB baseline.

A real-time defect detection system deployed on NVIDIA Jetson Nano for construction and manufacturing quality inspection. YOLOv5 was trained on a custom 1,300-image dataset annotated via LabelImg and augmented via Roboflow, covering cracks, bubbles, dents, and scratches. An Intel RealSense RGB-D camera provides depth data enabling 3D defect localization — distinguishing shallow surface-level cosmetic defects from deeper structural ones. The full pipeline runs inference on the Jetson Nano at real-time speeds.