Skip to content

3D Object Detection & Tracking from LiDAR Point Clouds for Self-Driving Car

License

Notifications You must be signed in to change notification settings

sai-aneesh/3D-Object-Detection-and-Tracking

Repository files navigation

SDCND : Sensor Fusion and Tracking

Overview

Detect objects in Lidar point-cloud data from the Waymo Open Dataset. Perform fusion between Lidar and camera detections and track objects using an Extended Kalman Filter. Implement data association and track management for the fusion solution.

The project consists of two main parts:

  1. Object detection: Extracting Lidar point-clouds from the Waymo data set, visualization, converting to a birds-eye view representation and executing a pre-trained NN on this data (FPN ResNet). Integrated a second NN and calculated standard evaluation metrics to compare their performance.
  2. Object tracking : An Extended Kalman Filter is used to track objects from both Lidar and camera detections. Data association based on the Single Nearest Neighbour method, track scoring and track management (initialization, confirmation, deletion), gating and field-of-view evaluation.

Answers to the project questions are in writeup.md

Stucture of the Project: Flowchart

Solution - Lidar-Camera fusion and object tracking

Final tracking output, using both Lidar and camera detections, EKF filter and track management: Fusion output

Solution - Lidar object detection

Object detection

Height and intensity extracted from Lidar range image: Range image

Point cloud visualization using Open3D Point cloud

3-channel BEV map containing density (red), height (green) and intensity (blue) data. BEV map

Integrated pre-trained FPN ResNet model for object detection ResNet detection

Precision, recall and metrics calculated using Darknet and FPN ResNet for detection Metrics using Darknet

Metrics using Resnet

About

3D Object Detection & Tracking from LiDAR Point Clouds for Self-Driving Car

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published