Implementation of birds detection system using the YOLOv5 model

This work considers a bird detection system using the YOLOv5 model in handling the challenge of avian-induced crop damages in agriculture. In this study, machine learning and mechatronics will be combined to develop an effective adaptive bird deterrent system. The methodology of the study will comprise data collection from farms and Kaggle, data augmentation techniques, and implementation on Google Colab. The baseline models developed with primary and secondary datasets are compared with a hybrid model and principal model on an augmented dataset containing 14,500 samples. Results show the impact of both dataset size and augmentation techniques is on model performance. The Hybrid Model returned a true positive rate for bird detection as high as 89%, while the principal model had further improvements, returning an accuracy of 85% in bird-type classification. The research proves that augmentation, diversity of datasets, and threshold optimization are three relevant areas related to the robustness of models for bird detection. Future work could be done by increasing dataset size, fine-tuning the model, and fusing multiple modalities to enhance real-world performance.