lndpd

Introduction

This serves as the official repository for the landing pad project. Dubbed "lndpd" this project aims to develop a camera integrated landing pad detection system for a drone. Currently, deep learning is being investigated as the intelligence behind the detection system. Pytorch is the deep learning library of choice preforming all AI task for this project. The overarching goal of this project is to locate the center of the landing pad and provide that information to flight controller. A sample image is provided below. The landing pad in this case is the traditional "H" helicopter landing pad.

The yellow dot represents predicted center of the landing pad. This image has been taken directly from the post processor revealing what the neural network sees. At a distance of 3 feet the deep learning algorithm can predict the center to with in an accuracy of 3 inches.

Basic Overview

The directory structure is as follows:

1. datagen: holds all code related to generating training data and validation data for the CNN.

   • The code in datagen references several external directories that hold the training and validation data. Additionally, some of the code requires a file at the top level of datagen directory for the "real data" labels.

2. pytorch: holds all code related to training and inferencing the CNN

   • The code in pytorch is self-contained and requires no external directories.
   • There exist two directories in pytorch named inference and train. The inference directory holds code related to inferencing the CNN and the directory train holds code related to training the CNN.

Pipeline

The pipeline for this project is setup in three phases: data generation, training, and inferencing. Data in the form of images and labels are created in the data generation phase. These images with the corresponding labels are then passed to the training phase where the CNN is trained and a saved model is produced. The inferencing phase utilizes the saved model to predict the newly presented images. The pipeline is setup this way to allow for independence between each phase. Thus, any phase can be modified will maintaining interoperability between the other phases.

Getting Started

A more detailed technical guide is provided as "tch_doc.pptx" in a PowerPoint format. However, a novice user looking to directly run this neural network would be most interested in the following lines of code:

  X = get_image()
  y = infer(X)

This code segment can be located on lines 64 and 65 in "inf.py" in the inference directory. An identical code segment can be found in "inf_cude.py" for cuda supported devices. In either case the function "get_image" will get an image from the attached webcam while "infer" will query the neural network for the predicted center coordinates. Thus, the output from "infer" is a four element vector arranged as

[x, y, w, h]

where x and y are the center coordinators as measured from the bottom left of the image. The output from "infer" can be utilized in anyway the user desires. As the program is written now the output will be written to the standard out. A previously trained model has been uploaded to this repository in the "train" directory. This model is trained over the traditional "H" landing pad similar to the sample image provided above. In addition to the code provided several python libraries will be needed to run the program. The libraries are as follows:

1. torch
2. cv2
3. numpy

Installing these libraries is as simple as a "pip3 install"