Use of deep learning algorithms for real-time detection of vessels in confined spaces using the Tensorflow framework

Abstract

Over 4515 small boat accidents were registered in the United State of America in 2012, resulting in 651 causalities and 22% of the accidents took place between two boats. It is, therefore, one of the most interesting applications for image analysis and recognition using deep learning, collision avoidance in passenger boats. Advances in parallel computing, graphic processing unit technology and deep learning have facilitated real-time image processing. The main objective of this study was to compare the performance metrics for different deep learning algorithms using pre-trained data sets. The algorithms used were: faster region-based convolutional neural networks, region-based fully convolutional network, and single shot multibox detector using the feature extractors: residual neural network, inception and convolutional neural networks for mobile vision applications to detect generic boats in confined waterways. These models were coded in Python programming language, using the framework Tensorflow and OpenCV library for image processing. The algorithms were pre-trained using the free images database posted on the web, Microsoft COCO. The use of these pre-trained models allowed making use of computers without graphic processing unit. As a result, it was found that the faster region-based convolutional neural networks and region-based fully convolutional network method compared to the single shot multibox detector method offer a small advantage precision if speed detection is not required, but the single shot multibox detector method is useful for case detectors in real time, however it did not perform as accurate when detecting small objects.