Research On Machine Vision-based Navigation Technology Of Automatic Guided Vehicle For Indoor Water Quality Testing

With the development of aquaculture to an intensive and high-yield model,indoor aquaculture has gradually become an important source of aquatic products by virtue of its short cycle and high yield.The Indoor high-density breeding method increases water pollution.Water pollution affects the survival rate of aquatic products and endangers the health of consumer.Water quality testing for indoor aquaculture ponds is essential.Compared with manual and wireless sensor networks,autonomous cruise vehicles can automatically detect all aquaculture ponds with only one set of equipment.Water quality detection by cruise vehicles reduces equipment costs and improves detection efficiency.As the key to autonomous cruising vehicles,navigation is a guarantee for reliable water quality testing.Considering that visual navigation is rich in information collected and suitable for various complex environments,this article uses machine vision to realize navigation of cruise car and conducts research on the key visual navigation technology.The main contents are as follows:First of all,by analyzing the navigation environment and requirements,formulate a technical plan,which is suitable for this subject.According to specific needs and plans,design the cruise vehicle hardware system,and calibrate the important camera modules to ensure that reliable navigation images can be collected.Then the environment of aquaculture sites is special,and some areas have dark and uneven lighting problems,which affects normal visual navigation.Aimed at the problem of large calculation volume and slow running speed of traditional MSR algorithm,an improved fast image enhancement algorithm is proposed.The image is quickly transformed into the HSV space.The V channel is decomposed by Haar wavelet and the mean filter is used instead of the Gaussian filter to complete the fast image enhancement.Experiments have shown that the improved algorithm is faster than traditional MSR and other fast enhancement algorithms,and can achieve better enhancement effects in low-light environments.Then it is difficult to detect obstacles quickly and accurately by deep convolutional networks on embedded devices.In order to solve the problem,the traditional SSD detection network is improved.This article uses the lightweight network Mobilenet V3 to replace the original backbone network VGG16 and deletes some non-key detection features.The loss function of network is optimized for higher detection accuracy.The data set of obstacle in the breeding base is made and it is used to train the improved network.Finally the obtained network runs at 15.6FPS on the NVIDIA development board.The speed of the network is 44.4%faster than that of Mobilenet V2-SSD,and its recognition accuracy is also slightly improved.Finally,in consideration of the environmental characteristics of aquaculture ponds,the gray value of the junction between the wall of aquaculture ponds and the ground is used to segment the navigation area.The center line of the area is fitted to extract the navigation direction angle.The global positioning of the cruise vehicle is completed by identifying the number of aquaculture ponds.The fitted monocular vision model is used to calculate the distance between the vehicle and the marking of aquaculture ponds to realize local positioning during the navigation process.Finally,through experiments,it is concluded that the error of the navigation direction angle extraction is within 2°,and the accuracy of the visual positioning is less than 8 cm,which meets the actual navigation requirements.