top of page

Wireless Communication and CNN-based object detection using UAV and Raspberry Pi

This project was done by me as part of a team of 4 individuals. The objective was to build an autonomous drone connected wirelessly to a base station using a Raspberry Pi microcontroller onboard. The primary aim was to use the Pi Camera on the drone to identify a retroreflective search target while flying utilizing the TensorFlow Lite framework. When the target is detected, the drone stops over the target and relays the location to the base station.

dronearch.png

I particularly worked on the development and deployment of the CNN architecture on the Raspberry Pi. Below is a brief overview of the same

The first step in developing the model was creating a Training Data Set. The algorithm learns various attributes of the target through these training data. Below are some of the key steps which we targeted for our training data set creation

  • Taking multiple aerial shots from different orientations and lighting conditions over the course of a day.

  • Image augmentation - Artificially creating training images through different ways of processing or a combination of multiple processing, such as random rotation, shifts, shear, and flips.

  • Image Labeling - Bound and label the target object in each training image. To accomplish this, we will be using tools such as ‘labelimg’ to automate the process.

​

algo.png

Flowchart Depicting Process Flow Being Followed for Development of CNN Model for Object Detection

We then used TensorFlow model maker to make use of the existing EfficientDetLite1 framework trained on the COCO dataset and repurposed it for our application by training it on our dataset.

Below are some of the parameter data used for training the model

  • Algorithm - Efficient Det Lite 1

  • Training Data Set size - 200 Images (180 used for training, 20 for validation)

  • Epochs - 50 (number of times training data was run through the algorithm)

  • Confidence threshold - 0.5 (can be increased or decreased based on results)

The model developed was able to achieve an average precision score of 0.837 which is considered very good for edge device application

​

Target being detected in test image with 0.94 confidence level

bottom of page