Underwater Object Detection From Videos

21st century is the century of ocean. The United Nations' "Agenda 21" suggests that the ocean is an essential component of a global life support system and a valuable asset for achieving sustainable development. Studying marine organisms plays a very important role in human development. It makes ocean study difficult that the marine and the underwater world are unfathomable. At present, people's understanding of marine life is almost through videos taken by underwater cameras fixed on devices such as autonomous underwater vehicles instead of fishing nets. It can cost several months to several years to manually process these videos. Therefore, the detection, tracking and classification of underwater creatures is a current hot spot.Moving targets detection has an important application potential in robot control, autonomous vehicle navigation, man-machine interface, medical imaging and video monitoring. With advances in technology, a variety of new technologies have been applied to target detection in more complex environments. But so far, no algorithm is applicable to all cases. Especially in the case of camera motion, no uniform method can tackle all kinds of camera motion. In addition, not only target motion, but also background motion caused by camera movement can lead to difficulty to extract the foreground from the background. However target detection is the basis of target tracking and classification, and detection results will have a great impact on objective understanding and analysis in the following-up processing. So it is a great challenge to work out a robust, precise and high-performance moving targets detection algorithm.Jellyfish in underwater videos are studied. It is a hard problem because they are small, with inconsistent brightness from the head to the tail, and captured by moving camera. The human eye can even hardly distinguish them from the sea water.Based on reviewing the previous research, and doing research on a number of detection algorithms and their applications, we proposes two new detection algorithms to deal with detection failure or fault detection because good results cannot be obtained by only using conventional methods.Firstly, the principle of intensity channel, color channel and moment channel which constitute the saliency map is analyzed in-depth, and then it is applied in target detection with central moment instead of color. Thus it turns out a new target detection based on background subtraction and saliency map. To model the background, the average of some frames before the current frame is calculated. The whole stable object detection system is established, where all the maps are obtained and the priori maximum number of objects in a single frame determines whether the target appeared. Experiments show that the algorithm can detect the target that is not so clear. That provides convenience for further processing such as tracking and classification.Because the video is captured while the camera is moving ahead, a special point called Focus of Expansion (FOE) exists. Another new object detection algorithm based on K-FOE residual map and ring segmentation is put forward. A Kalman filter is applied to predict the accurate coordinate of FOE to calculate the residual map. And then a different threshold is set according to the different distance between the object center and FOE and is used to update the object template.