Action Recognition From Thermal Videos Using Joint and Skeleton Information

Although various studies based on thermal images have been conducted, few focused the simultaneous extraction of joints and skeleton information an object from a image, performed human action recognition using this information. Unlike in case visible light images, performing joint detection generation often leads to complete disappearance spatial such as joints. In case, it is extremely difficult extract object. Moreover, accuracy significantly reduced owing issue. Therefore, new method proposed study address these issues. method, original 1-channel image was converted into 3-channel then were combined improve performance. A generative adversarial network (GAN) used for extracting addition, research recognize actions conducted extracted by method. The combining convolutional neural (CNN) long short-term memory (LSTM). As result experiments self-collected open data, found that shows good performance compared other state-of-the-art methods.