Autonomous Robot for Food Delivery

The Autonomous Drone for Urban Food Delivery project aims to provide efficient, contactless food delivery in city environments. With sensors including camera, LiDAR, GPS, IMU, and ultrasonic for obstacle detection, the drone can operate in both manual and autonomous modes. Utilizing an in-house autonomous stack, the drone combines LiDAR, GPS, and IMU data to navigate urban landscapes precisely, avoiding obstacles and ensuring timely deliveries. This technology supports sustainable, quick, and convenient food delivery directly to customers in densely populated areas.

Counting Cars in a Multistorey Car park

This project focuses on automating car counting in a parking lot using a camera installed at the entrance/exit. By replacing manual inspection, the system efficiently tracks cars entering and exiting the lot, providing real-time updates on available parking spots for drivers. A lightweight object detection model was developed to run on the Raspberry Pi CPU, eliminating the need for expensive devices like Nvidia Xavier. The model achieved an impressive 17 FPS on a Raspberry Pi 3, ensuring cost-effectiveness and high performance. Additionally, a tracker was implemented to distinguish whether a car is entering or exiting, avoiding the need for an additional camera installation and further optimizing the solution. The collected numbers and statistics were transmitted to a database, enabling seamless integration with a website and app to display parking availability to users in real time.

Building footprint extraction from high resolution satellite images

My Master's thesis focuses on pixel-wise image segmentation to automate tasks such as building detection from 3-band monocular satellite images using convolutional neural networks (CNNs) like FCN, SegNet, UNet, Deep UNet, PSPNet and Mask RCNN. Manual extraction of buildings, roads, and other features from satellite imagery is a time-intensive process, especially in rapidly growing cities. By leveraging deep learning, computational advancements, and high-quality data acquisition, this project aims to streamline this process. The core objective is to evaluate the transferability of trained neural networks to detect buildings in unseen geographical regions, addressing variations in building attributes across different locations. The research also explores factors such as image resolution, quality, building separation, and texture of non-building areas. Transferability is assessed using metrics like completeness, correctness, overlap with ground truth data, and accuracy. This work contributes towards creating a generalized solution for building detection tasks in diverse environments.

LIDAR Point Cloud Processing

Real-time processing of raw point clouds from various lidar types, including 2D single-line lidar (RP Lidar), 3D solid-state lidar (Livox, AEye, and Innoviz), and 3D mechanical lidar (Ouster, and Velodyne). The key functionalities include:

• Compression of point clouds for efficient streaming.
• Feature extraction and obstacle detection.
• Identification of regions of interest (ROI) and non-interest.
• Real-time point cloud registration on embedded devices.
• Live streaming of processed point clouds using a cloud server for edge devices.

The entire workflow was designed for seamless operation in real-time, leveraging embedded platforms to enable effective lidar data utilization in applications like navigation, mapping, and obstacle avoidance.

Simultaneous Localization and Mapping (SLAM)

This project focuses on implementing a real-time Simultaneous Localization and Mapping (SLAM) system with millimeter-level precision. The system is designed to operate efficiently on embedded devices and is robust for both constrained and unconstrained environments. Key features include:

• Point cloud-based graph SLAM for accurate mapping.
• Feature-based loop closure to enhance map consistency.
• Integration of GPS, IMU, and gyroscope data for additional assistance, when required.

This SLAM solution is optimized for resource-constrained platforms and is ideal for applications requiring precise localization and mapping in diverse and dynamic scenarios.

Path planning

This project involves designing a grid-based path planner for autonomous navigation in urban streets and warehouses. The planner incorporates algorithms such as waypoints, Dijkstra's, and A* to ensure efficient route generation. Developed for food delivery robots and forklifts, the system includes real-time obstacle avoidance and dynamic re-routing capabilities to handle changing environments seamlessly. This solution enables safe and efficient navigation for autonomous vehicles in complex, dynamic settings.

Remote Operation

This project enables remote operation of robots and forklifts, allowing users to control vehicles from a distant location instead of physically driving them. The system supports any type of controller and uses WebRTC for seamless communication. Control commands are sent to a server, and vehicles subscribe to these commands for execution. The user interface provides real-time statistics of the vehicle, while the edge system integrates cameras for visual inspection and lidar, radar, or ultrasonic sensors for obstacle detection. This solution ensures safe and efficient remote operation in diverse environments.

Free Parking Space Detection in a Parking Lot

This project focuses on automating car counting in a parking lot using a camera installed at the parking lots. By replacing manual inspection, the system efficiently tracks each parking space , providing real-time updates on available parking spots for drivers. A lightweight object detection model was developed to run on the Raspberry Pi CPU, eliminating the need for expensive devices like Nvidia Xavier. The model achieved an impressive 17 FPS on a Raspberry Pi 3, ensuring cost-effectiveness and high performance. The collected numbers and statistics were transmitted to a database, enabling seamless integration with a website and app to display parking availability to users in real time.