AbstractThis master thesis proposes an introduction to Region-Based CNN with the use of Faster RCNNs via the reproduction of a scientific protocol  in the context of precision agriculture. Two Faster RCNNs embedded on swarm MAV have been developed to detect weeds in sugar beet crops simulated on Unity3D. These Faster RCNNs are of different depths. The objective is to increase the performance of the shallowest Faster RCNN to optimize computational resource consumption. The detections from previous passes of a MAV over the same area will be used to improve the detections of this Faster RCNN. The performances are compared. The past detections have no impact on the detection performances of the shallowest Faster RCNN.
A confusion is made between epochs and iterations during the development and the training of the Faster RCNNs. The TorchVision library is used in first time. Then it is replaced by the maskrcnn-benchmark framework. The dataset used has a lack of images. As a result, it was not easy to train these networks. The results obtained in the framework and those obtained in the article defining the methodology  will be compared.
Finally, this work will propose different ways of improvements.
Federico Magistri, Daniele Nardi, and Vito Trianni. Using prior information to improve crop/weed classification by mav swarms. In 2019 IMAV/ 11th INTERNATIONAL MICRO AIR VEHICLE COMPETITION AND CONFERENCE(IMAV), pages 67–75, 2019.
|Date of Award||23 Jun 2021|
|Supervisor||Elio Tuci (Supervisor) & Guillaume Maitre (Co-Supervisor)|
- Artificial intelligence
- Computer Vision
- Object detection
- Faster RCNN
- Precision agriculture
- Weed control