This repository demonstrates the implementation of multi-sensor tracking and state estimation algorithms for autonomous systems, focusing on radar and lidar fusion using Extended and Unscented Kalman Filters.
The work reflects a system-level approach to tracking, including measurement modeling, nonlinear estimation, and integration into a broader tracking pipeline, with future extensions toward multi-object tracking and environment mapping.
This portfolio demonstrates sensor fusion across different levels:
-
EKF (Extended Kalman Filter)
Dataset-driven estimation using linearization (Jacobian) for nonlinear radar measurements. Focus on correctness and RMSE evaluation. -
UKF (Unscented Kalman Filter)
Nonlinear estimation using sigma points with a PCL-based simulation environment. Focus on system behavior and visualization. -
Radar-Camera Fusion
Real-time multi-sensor integration combining perception and tracking.
This progression reflects increasing system complexity: EKF → UKF → Multi-sensor fusion system.
- Designed as part of a broader tracking pipeline: sensor → measurement → data association → state estimation (EKF/UKF) → track management
- Measurements from radar/lidar are associated to tracked objects using gating and nearest-neighbor concepts
- State estimation is performed using nonlinear filters (EKF/UKF) for robust tracking under sensor noise
- Track lifecycle management includes initialization, update, and deletion based on measurement consistency
- Designed to scale toward multi-object tracking and integration with mapping modules (e.g., occupancy grids)
- Measurements are associated to existing tracks using gating (Mahalanobis distance) to reject unlikely matches
- Nearest-neighbor approach is used to assign measurements to predicted states
- Handles uncertainty in sensor measurements and avoids incorrect track updates
- Designed to extend toward probabilistic methods (e.g., JPDA) for multi-object tracking scenarios
Tracking Pipeline Overview:
Radar / Lidar → Measurement Processing → Data Association → EKF / UKF → Track Management → (Future: Occupancy Grid Mapping)
A real-time C++ prototype for multi-object tracking and sensor fusion using camera detection, Kalman filtering, and EKF-based radar updates.
- Real-time multi-object tracking in C++
- Kalman filter (constant velocity model)
- EKF radar update (range + angle)
- Camera-radar association
- Detection filtering and tracking stabilization
- Used motion detection for real-time performance on CPU
- Integrated YOLO as optional detector (higher accuracy, lower FPS)
- Used Kalman filter to maintain tracking between detections
- Motion-based object detection (real-time)
- Multi-object tracking with persistent IDs
- Constant-velocity Kalman filter
- EKF radar update (range + angle)
- Simulated radar measurements
- Optional YOLO detector (OpenCV DNN)
C++, OpenCV, CMake, Visual Studio
🔗 Repo: https://github.com/Vaiy108/radar-camera-fusion
This project builds the theoretical foundation for the real-time radar-camera fusion system implemented above.
🔗 Project link: unscented-kalman-filter
Key concepts:
- Nonlinear state estimation
- Sigma points (unscented transform)
- Radar + lidar fusion
Implemented an Unscented Kalman Filter (UKF) to fuse radar and lidar measurements for tracking a moving object under nonlinear motion (CTRV model).
The UKF improves over EKF by avoiding linearization and instead propagating sigma points through nonlinear dynamics for better accuracy.
- Full UKF pipeline (initialization, prediction, update)
- Sigma point generation and propagation
- CTRV motion model for nonlinear dynamics
- Radar update (nonlinear measurement model)
- Lidar update (linear measurement model)
- Angle normalization and numerical stability handling
State vector includes:
- position (
$$p_x, p_y$$ ) - velocity (v)
- yaw angle
$$(\psi)$$ - yaw rate
$$(\dot{\psi})$$
Motion model:
CTRV (Constant Turn Rate and Velocity)
Key components:
- Augmented state with process noise
- Sigma point generation using unscented transform
- Nonlinear prediction step
- Sensor-specific update steps:
- Linear update (lidar)
- Nonlinear update (radar)
- Full UKF pipeline:
- Initialization from first measurement
- Time-based prediction
- Sensor fusion updates
- Angle normalization for stability
- Numerical safeguards (division-by-zero handling)
- Clean modular C++ design using Eigen
- Demonstrates nonlinear sensor fusion (UKF)
- Shows probabilistic state estimation in C++
- Theoretical Foundation for real-world tracking systems
Sensor fusion using Jacobian-based linearization for radar and lidar measurements with dataset-driven evaluation. This project complements the UKF implementation by providing a baseline linearization-based approach for comparison.
Features:
- EKF prediction and update pipeline
- Radar nonlinear measurement handling using Jacobian
- Lidar linear update
- RMSE-based evaluation using ground-truth dataset
- C++ implementation with Eigen
- CMake build system
🔗 Project link: extended-kalman-filter
- C++
- Eigen (linear algebra)
- CMake
- Git/GitHub
- C++
- Sensor Fusion
- Kalman Filtering (EKF / UKF)
- Nonlinear State Estimation
- Unscented Transform
- Radar & Lidar Fusion
- Linear Algebra (Eigen)
- System Modeling (CTRV)
Vasan Iyer
Sensor Fusion / Autonomous Systems Engineer
Focus areas:
- Sensor fusion & state estimation
- Autonomous systems
- Flight dynamics & control
- Embedded systems (C++, Python)
- UAV systems & simulation
GitHub: https://github.com/Vaiy108