The Design And Implementation Of Sea Cucumber Target Recognition System Based On YOLO-v3

The technology of automatic intelligent farming and fishing which applied to aquaculture is developing continuously with the development of China’s agricultural modernization.Currently the main fishing method is still manual diving fishing in the sea cucumber aquaculture cofferdam in the Bohai Bay area,fishers must dive into the cofferdam to capture sea cucumbers with oxygen cylinder or other oxygen supply equipment on their back.This is a dangerous fishing way with high cost especially fishing in the deep water which is considered as a highrisk task in the industry.It has a huge market and broad prospects to develop an intelligent sea cucumber fishing robot which can capture sea cucumber automatically to substitute manual diving fishing.Firstly,the problem of underwater detection and positioning for the sea cucumber targets should be resolved.Therefore,a sea cucumber objection detection and positioning system based on YOLO-v3 has been proposed and implemented in this article,this system can detect sea cucumber and send the target’s position information to the intelligent sea cucumber fishing robot.There is the main work in this article:Firstly,analyze the precision and real-time requirements of the grasping action of the intelligent sea cucumber fishing robot,and overall design the sea cucumber objection detection and positioning system according to the requirements which analyzed before,and divided the recognition system into sea cucumber target detection modules and target positioning modules with different functions,which respectively complete the two functions of target detection and target positioning.Secondly,analyze the image preprocessing and data set expansion methods which can be suitable for underwater images according to the specificity of underwater images.The combined ACE and defogging algorithm is used to enhance underwater images,and the image enhancement and the image vertical flip methods are used to expand the underwater sea cucumber image dataset.Thirdly,analyze the advantages and disadvantages of the convolutional neural networks and the YOLO series networks,then choose the YOLO-v3 network as the basic network for targets detection according to the requirements of the intelligent sea cucumber fishing robot,and adjust the network’s parameters based on the sea cucumber image dataset which is built on the video captured in the real sea cucumber cofferdam.Train the network on this dataset and compare other YOLO network’s performance.Then implement the detection module on the ROS platform.Finally,design the target positioning module combined with the underwater stereo camera’s depth measurement and positioning functions,including the correction of both depth measurement and positioning,then implement the positioning module on the ROS platform with the correction data.Then test the whole system.