Abstract

This project focuses on developing an automated system for detecting and counting rabbits in a rabbit farm using deep learning techniques. The primary objective is to create a robust and accurate model capable of identifying individual rabbits from video footage and counting the total number of rabbits in each frame. An additional feature of the system is to track rabbits moving between the inside and outside sections of the farm, separated by a door barrier, and count their movement in real-time. This system aims to aid farm management by providing precise, continuous monitoring and data collection to enhance operational efficiency.

The project employed a TensorFlow Lite-based object detection model for real-time rabbit detection and counting. Key steps included extensive data collection, manual annotation of frames using LabelImg, model training, evaluation, and deployment. Videos recorded by CCTV cameras in the rabbit farm were processed to extract frames, which were then annotated to create a high-quality labeled dataset. Initial experiments involved using YOLOv5, but after extensive testing and iterative refinements, TensorFlow Lite was chosen for its lightweight and efficient performance.

The trained model was successfully tested across multiple platforms, including personal devices (MacBook Pro 2020), university virtual machines, and Google Colab Pro for optimal resource utilization. Despite challenges like computational limitations, compatibility issues, and the need for efficient app deployment, the final system achieved significant milestones. The system was delivered as both a standalone application for macOS and Windows, built with PyInstaller, enabling seamless usage by non-technical users.

The model is now deployed in real-world scenarios, currently being utilized by the Agriculture Department of our university to monitor rabbit populations and movement in real time. The system has demonstrated high accuracy and practicality, overcoming challenges such as handling large datasets, platform dependencies, and the creation of deployable apps.

In conclusion, this project not only demonstrates the feasibility of deep learning in automated animal detection and counting but also highlights the potential for scaling such systems to other agricultural applications. Future work will focus on further improving model accuracy, integrating additional features like anomaly detection, and transitioning towards an open-source framework with detailed documentation and an accessible GitHub repository for global adoption.