SEPPI Camera Traps
The camera traps used in the SEPPI project aim to automatically capture pollinator visitation on real flowers. An evolution of the Insect Detect DIY camera trap for monitoring pollinators on artificial flower platforms, specifically optimized for natural flower environments.
To enable the monitoring of plant-pollinator interactions, several aspects of the existing cameras required adaptation. Monitoring insect visitors on natural flowers presents unique challenges: detecting insects against heterogeneous, non-flat backgrounds, configuring optimal focus and lighting systems, and extending the sampling duration for individual flowers.
Key adjustments:
To enable the monitoring of plant-pollinator interactions, several aspects of the existing cameras required adaptation. Monitoring insect visitors on natural flowers presents unique challenges: detecting insects against heterogeneous, non-flat backgrounds, configuring optimal focus and lighting systems, and extending the sampling duration for individual flowers.
Key adjustments:
- Insect detection on natural flower backgrounds
- Manual configuration of focus and lighting in the field
- Individual operational duration per flower
How it works
Field Deployment and Setup
The camera traps monitor individual flowers in their natural habitat. Each camera is fixed in the ground near a selected flower, with the enclosure height adjusted for optimal viewing distance.
Configuration and Control
An integrated web app enables real-time field configuration of camera settings. This allows manual focus adjustment and configuration of other parameters essential for monitoring flower visitation against diverse, non-flat, and heterogeneous natural backgrounds.
Recording Process
Once configured, the system begins automated recording using a dual-stream approach. The OAK-1 camera runs simultaneous low-quality (320x320px) and high-quality (3840x2160px) streams through the DepthAI Python API.
The low-quality stream feeds a YOLO detection model for rapid insect identification, while detections are synchronized with high-quality frames on-device. This dual-stream method maximizes both inference speed and tracking accuracy. Detected insects are automatically cropped from high-quality frames and saved to an SD card with complete metadata.
The low-quality stream feeds a YOLO detection model for rapid insect identification, while detections are synchronized with high-quality frames on-device. This dual-stream method maximizes both inference speed and tracking accuracy. Detected insects are automatically cropped from high-quality frames and saved to an SD card with complete metadata.
Classification and post-processing
The cropped images of detected insects will subsequently be classified into taxonomic groups or classes, for example using the open-source foundation model BioCLIP 2. A dedicated post-processing pipeline will support this classification, filter out false detections and repeated recordings of the same tracked pollinator individual, and generate a final interaction matrix. This matrix can then be used to analyse plant–pollinator network structure and diversity.
BiodivMon Call
|
Duration of the project
01.04.2024 - 31.03.2027 |
|
