Real-Time Weed Detection using Machine Learning and Stereo-Vision

Weeds are a problem as they compete with desirable crops and use up water, nutrients, and space. Some weeds also get entangled in machinery and prevent efficient harvesting. Hence weed removal systems are a necessity. Development of a successful weed removal system involves correct identification of the unwanted vegetation. The paper proposes a Real-time weed detection system that uses machine learning to identify weeds in crops and stereo-vision for 3D crop reconstruction. Structure from motion technique is utilized on a video of a farm to generate a 3D point cloud. The machine learning model is trained on two manually created datasets of cucumber and Onion crop. Convolutional Neural Networks (CNN) and ResNet-50 algorithms are used to train the machine learning models. It is seen that the ResNet-50 model outperforms the Convolution Neural Networks model. ResNet-50 model gives an overall accuracy of 84.6% for the cucumber dataset while it gives an accuracy of 90% for the Onion crop dataset.