Evaluating the Effectiveness of an Object Detection Pipeline to Support Surveillance of Unintended Passage
Sethu Mettukulam Jagadeesan, Jordan Leh, Jonathan Gregory, Jesse Eickholt, Daniel Patrick Zielinski · 2024
This research paper presents a novel approach to fish surveillance, specifically detecting fish jumping out of water, by leveraging deep learning-based object detection techniques. The study focuses on the use of an EfficientDet-Lite model, a lightweight and efficient model suitable for edge computing devices. The model was trained using a machine learning pipeline that significantly reduces the amount of data requiring manual review. The performance of the model was evaluated using standard average precision (AP) metrics and manual image- and clip-based evaluation. For the TFLite model running on the GPU, the precision derived from the manual evaluation was 0.841 and the recall was 0.279. For the same model compiled for and run on the Coral Edge TPU, the precision was slightly lower at 0.838 with a recall of 0.274. In the clip-based evaluation, the Edge TPU in the Frigate environment achieved a precision of 0.030 and a recall of 0.11. Although the AP metrics appear relatively low, the model demonstrated a high capacity to accurately detect and approximate the location of fish, which is the primary objective of this research. The model was further evaluated in a simulated production environment, demonstrating its potential for real-time fish detection during jumping events. Overall, this study contributes to the field of fish surveillance by introducing an efficient and effective method for fish detection. This research serves as a proof of concept for fish emergence observation using object detection.