About the presenter
Boglárka Ecsedi
I am a senior high school student of the Istvan Bocskai Secondary Grammar School in Hajdúböszörmény, Hungary. Through collaborations with research groups at the University of California, Santa Cruz and the Medical University of Vienna, I have been doing scientific research for two years. I have developed machine learning and computer vision methods to solve problems ranging from detecting hazardous natural phenomena to analyzing heterogeneous tumors. I have a firm interest in studying Computer Science with a focus on Artificial Intelligence, Computer Vision and Robotics.
Personal video
Abstract
Rip Current Detection - An Orientation-aware Machine Learning Approach
An ever-changing hazardous natural phenomenon – called a rip current – causes numerous fatal accidents all over the world. To address this problem, I use an image processing algorithm to detect and localize rip currents with a powerful near real-time deep neural network called Faster R-CNN (Ren, He, Girshick, & Sun, 2016). I specially assembled training and validation datasets of more than 1000 images of rip currents and used transfer learning to improve the baseline model with my custom data. The results showed an accuracy of 85.19% for an IoU threshold of 0.5 and an average precision (AP) of 0.371. To further improve the performance of the detection model, I developed an orientation-aware region proposal layer and incorporated it to the framework of Faster R-CNN. This layer predicts rotated bounding boxes in contrast to the traditional axis-aligned bounding boxes. Based on evaluation using the IoU parameter, the findings revealed that the orientation-aware region proposal layer outperforms the axis-aligned region proposal layer of the original Faster R-CNN. In one of the experiments, I found an average efficiency improvement of 11.07%. The development resulted in detecting rip currents with a higher efficiency, allowing the algorithm to adapt to many angles, positions of the object and different perspectives. An automated rip current detection system using the improved detection algorithm is under development. My goal is to turn this algorithm into a mobile application. This approach contributes to the deeper understanding of rip currents, to the early identification of the hazard, thus preventing accidents and protecting human lives.
Keywords: deep neural network, object detection and localization, orientation-aware object detection, rip current