Hornet Detection & Tracking
Model Training & Evaluation · Team of 3
A Streamlit application that detects and tracks invasive hornets threatening bee populations, using YOLO based computer vision. It was a team project where I led model training and evaluation.
Overview
About the project
Invasive hornet species are a serious threat to bee populations and the wider ecosystem. Monitoring hives by hand does not scale, and early detection is critical, so the spotting itself needs to be automated.
Our three person team built a Streamlit application around YOLO based object detection. I led model training and evaluation while teammates handled data preprocessing. We annotated training data with Roboflow, trained and compared YOLOv7 and YOLOv8 models on Google Colab's GPUs, and evaluated detection and tracking performance on video footage.
The result is a working prototype that automates hornet detection and tracking in video, removing labour intensive manual review. It also provides a foundation that can be adapted to other ecological monitoring tasks.
System Overview
At a glance
A team computer vision project that automates the spotting and tracking of invasive hornets in video, removing the need for slow manual review. We trained and compared YOLOv7 and YOLOv8 detectors on a custom annotated dataset and wrapped the result in a Streamlit application. I led model training and evaluation in a three person team.
How It Works
System Architecture
Roboflow
Annotation of the hornet training data
YOLOv7 & YOLOv8
Detection models trained and compared
Google Colab
GPU environment used for training
Streamlit App
Interface for running detection on footage
Process Flow
How the project moves from start to finish, step by step.
Annotate
Label hornet footage as a training dataset in Roboflow.
Train
Train and compare YOLOv7 and YOLOv8 on Google Colab GPUs.
Track
Detect and follow hornets across video frames in the app.
System Breakdown
Experiment Setup
The work was structured so the model choice would be defensible, not arbitrary.
Dataset
Roboflow Annotation
Hornet footage labelled directly in Roboflow.
Frame Sampling
Frames extracted from video for annotation.
Train and Validation Splits
Honest splits so results are comparable across models.
Models and Training
YOLOv7
Trained as a comparison candidate.
YOLOv8
Trained on the same dataset with matching hyperparameters.
Google Colab GPUs
GPU environment that kept training cycles short.
Application
Streamlit App
Loads the trained model and runs detection on uploaded footage.
Tracking Layer
Follows each detected hornet across consecutive frames.
Workflows
Project Phases
Three phases, with model comparison treated as a first class output.
Annotate
Purpose: Build a dataset both models can be trained on fairly.
How it works
- Extract frames from hornet video footage.
- Annotate every hornet in Roboflow with consistent labelling rules.
- Split the dataset into train and validation sets.
Train and Compare
Purpose: Pick the right model with real numbers behind the decision.
How it works
- Train YOLOv7 and YOLOv8 on the same data with matching settings.
- Evaluate each on the validation set.
- Compare detection performance side by side.
Detect and Track
Purpose: Make the chosen model useful in the field.
How it works
- Run inference on video footage in the Streamlit app.
- Track every detection across consecutive frames.
- Produce annotated output a beekeeper can scan quickly.
What It Does
Features & Capabilities
Automated Detection
Spots hornets without slow, manual video review.
Model Comparison
YOLOv7 and YOLOv8 evaluated side by side for the task.
Object Tracking
Follows each detected hornet across video frames.
Scalable Use
Adaptable to other ecological monitoring tasks.
Running It