CTDR-Net: Channel-Time Dense Residual Network for Detecting Crew Overboard Behavior

Abstract

The efficient detection of crew overboard behavior has become an important element in enhancing the ability to respond to marine disasters. It remains challenging due to (1) the lack of effective features making feature extraction difficult and recognition accuracy low and (2) the insufficient computing power resulting in the poor real-time performance of existing algorithms. In this paper, we propose a Channel-Time Dense Residual Network (CTDR-Net) for detecting crew overboard behavior, including a Dense Residual Network (DR-Net) and a Channel-Time Attention Mechanism (CTAM). The DR-Net is proposed to extract features, which employs the convolutional splitting method to improve the extraction ability of sparse features and reduce the number of network parameters. The CTAM is used to enhance the expression ability of channel feature information, and can increase the accuracy of behavior detection more effectively. We use the LeakyReLU activation function to improve the nonlinear modeling ability of the network, which can further enhance the network’s generalization ability. The experiments show that our method has an accuracy of 96.9%, striking a good balance between accuracy and real-time performance.