« back to Publications
Laradji, I., Schmidt, M., vanPoorten, B., Godin, T. and Greenberg, S. (2017)
Using Deep Learning to Automate Boat Detection for Fishing Effort Estimates. In Proceedings of the 8th World Recreational Fishing Conference – WRFC8. (Victoria, Canada), July 16-20.
Type: Refereed conference (abstract and poster).
Note: The PDF contains the abstract and the poster presented at the conference.
Download as: Pdf
Abstract.
Estimating angling efforts is an important statistical measure for managing recreational fishing on lakes. One common strategy is to capture lake images over long periods of time and count the number of anglers. However, analyzing lake images takes a vast amount of human effort and time. We are developing a Computer Vision algorithm that attempts to perform these counts automatically, and here we report on our progress to date. Using fully-convolutional neural networks (with the help of data augmentation) the system learns and then determines the regions where boats are located in the lake images. When comparing the algorithmic results with manual counting, our results over a single set of lake images were reasonably accurate: about 10% false positive rate and false negative rate were seen. Our work to date has exposed many challenges that remain to be overcome, but the progress is promising. Our future work will test the algorithm on a variety of data sets and will analyze images for shore anglers as well as boats. If successful, we expect automation to significantly reduce the workload on human experts, where their job is to confirm or reject predictions rather than laboriously search for and count anglers and boats in lake images.
