Estimate sequential poses for wireless endoscopic capsule based on encoder-decoder convolutional neural network toward autonomous navigation

Abstract Capsule endoscopy is useful for the detection of a wide range of digestive diseases. Further estimation of the capsule's position, orientation, namely pose for short, allows more intuitive navigation and correct diagnosis with patient-specific capsule localization and image mapping onto a global coordinate. There are traditional methods for pose estimation like visual odometry, but its readings can have increasing drift overtime. The new advanced research in deep learning gives us better estimation accuracy without the use of costly sensors and equipment. Our model utilizes an encoder-decoder convolutional neural network architecture. We can predict the subsequent position of the capsule in the digestive tract with excellent precision. The neural network furnishes 6 DOF values when fed with an RGB image from the camera of the capsule, without image preprocessing except for resizing and normalization. This can be practiced in realtime by reducing the inference time and increasing the frame rate per second.