Project DetailsThis project focuses on analyzing shark-human interactions using drone footage. Sharks are a major attraction along the coasts of Israel, especially near power plants where they aggregate during the winter season. The primary goal of the project was to develop an automated algorithm that identifies and classifies different behaviors of sharks interacting with humans using computer vision and machine learning technologies.
The developed algorithm includes four main stages: detection and localization of sharks, tracking their movements throughout the video, and classifying their behavior. For the detection and localization phase, we utilized the YOLO architecture, while for tracking, we chose the deepSORT algorithm. In the behavior classification phase, we analyzed the distance between sharks and swimmers over time to classify behaviors as approaching, escaping, or neutral.
Detection and tracking results were limited due to data quality issues (unlabeled data) and model biases, which hindered the full success of the system. However, initial results showed the algorithm's ability to distinguish shark behaviors in specific situations, and we presented directions for future improvements.
