LIDAR · IMU · AI · 3D RECONSTRUCTION

TURNING A 2D LIDAR
INTO A 3D SCANNER

A servo motor and an IMU give a cheap 2D LiDAR a third dimension — tilting it through space, tracking every angle, and stitching the slices into a full 3D point cloud, at a fraction of the cost of commercial sensors.

2D → 3D
SENSOR CONVERSION
80
TRAINING EPOCHS
~$120
EST. HARDWARE COST
SCROLL TO EXPLORE
01ABOUT THE PROJECT

Giving a 2D sensor a third dimension

Most affordable LiDAR units only see in 2D — a single flat slice of the world, with no sense of height. Commercial sensors that scan in full 3D cost thousands of dollars, putting real-time 3D mapping out of reach for students, hobbyists, and small robotics teams. We set out to close that gap using parts that together cost a fraction of a single commercial sensor.

The core idea: take a cheap 2D LiDAR and physically tilt it with a servo motor while continuously scanning. An IMU tracks the exact angle of every tilt, so each 2D scan line can be placed correctly in 3D space. Stack enough of these oriented slices together, and a full 3D point cloud emerges from a sensor that was never designed to produce one.

That point cloud is then fed into our own custom-built AI model, trained on the JackRabbot Dataset and Benchmark (JRDB), to detect and classify objects in the scanned environment — turning raw geometry into scene understanding. One application we're particularly drawn to is disaster response, where a low-cost 3D map of a space can help before anyone walks in — but the same pipeline applies to indoor mapping, robotics, and general 3D reconstruction.

01Convert a low-cost 2D LiDAR into a 3D scanner
02Use a servo + IMU to track orientation per scan line
03Reconstruct a full 3D point cloud from stacked slices
04Integrate AI-based object detection on the point cloud
05Visualize point cloud data in real time
06Provide an affordable alternative to commercial 3D systems
02HARDWARE COMPONENTS

Four parts, one scanner

Every component was chosen to keep the bill of materials low without sacrificing the orientation accuracy the 2D-to-3D conversion depends on.

SENSOR

YDLIDAR X2

2D LiDAR providing the raw range scans that get stacked into a 3D point cloud.

COMPUTE

Raspberry Pi 5

Runs point cloud generation and hands data off to the AI detection pipeline.

CONTROLLER

ESP32

Drives the tilt sweep and synchronizes orientation data in real time.

DISPLAY

SmartElex 5" TFT LCD

Capacitive touch display for on-device control and live point cloud feedback.

03SYSTEM ARCHITECTURE

From raw scan to detected object

LiDAR X201
Raspberry Pi 502
Point Cloud Generation03
AI Detection Model04
3D Visualization05
ESP32 CONTROLS
  • Servo movement (tilt sweep)
  • IMU angle reading
  • Synchronizing each scan slice with its orientation
04AI INTEGRATION

Modified RPEA model, trained on JRDB

The generated point cloud is passed through a modified version of the Residual Path Network with Efficient Attention (RPEA) detection model with a multi-branch detection head to specialize in occlusion patterns and distance bands in order to better detect humans in 3D space — no camera required, which also helps in low-light or visually cluttered environments.

Modified RPEA
MODEL ARCHITECTURE
JRDB
TRAINING DATASET
80
EPOCHS TRAINED
PyTorch
FRAMEWORK
HUMAN
05TECHNOLOGY STACK

What runs underneath

AI / ML
  • Python
  • PyTorch
  • Modified RPEA
  • JRDB Dataset
HARDWARE
  • Raspberry Pi 5
  • ESP32
  • SmartElex 5" TFT LCD
  • YDLIDAR X2
VISUALIZATION
  • Open3D
  • Point Cloud Rendering
06TEAM

Who built it

AH

Aahana Hajariwala

FRONTEND DEVELOPER
KK

Karthik Kumar

FIRMWARE ENGINEER
RB

Rohan Alex Basil

BACKEND DEVELOPER
ST

Sparsh Tyagi

HARDWARE DESIGNER
07FUTURE SCOPE

Where this goes next

Real-time autonomous navigation
Indoor 3D mapping
Robotics and obstacle avoidance
Search-and-rescue applications
Smart city monitoring
Security and surveillance
Enhanced object classification