STAR: Sea Turtle Basic Activity Recognizer Network via Efficient Transformer

Sea turtles are endemic to several islands and are crucial in maintaining ecosystem balance. However, they face significant threats from poaching, extreme weather conditions, and abnormal activities. Monitoring and protecting their habitats has become a key focus in blue economy research, a priority program in several developed nations. Observing sea turtle behavior not only aids in conservation efforts but also supports the development of behavioral analysis that can serve as metrics for further research. In this work, we propose STAR, a basic turtle activity recognition system, employing an efficient deep learning approach. This system categorizes simple activities such as swimming, eating, hiding, resting, and other distinct behaviors. An effective network is offered by utilizing enhanced modules to discriminate activity features involving sequential frames. In addition, a lightweight transformer is presented as a novel attentive module that improves the feature extraction ability. Another contribution offers a new video dataset of sea turtle activity recognition captured underwater from various turtle poses. Several extensive experiments analyzed the performances of essential modules and achieved satisfactory results compared to other lightweight models. The efficiency of the STAR model shows that it obtained fast data processing speed without abundant computational resources in single-frame analysis. Applying this recognition system requires sea turtle detection to capture the area of interest in the beginning process.