Road Obstacle Recognition Technology For Self-driving Feeder Vehicles For Aquaculture

At present,the aquaculture industry is gradually developing in the direction of science and technology.The artificial breeding method has high labor cost and timeconsuming and labor-intensive.Its limitations force people to develop a time-saving and labor-saving self-driving feeding vehicle to replace the artificial breeding operation,The problem of safe driving of autonomous vehicle is a hot and difficult point in the field of autonomous driving.This paper proposes an obstacle recognition algorithm based on deep learning for obstacles on roads in aquaculture environment,and the structure of the YOLOv4 algorithm is improved,so that the algorithm can accurately and quickly identify the obstacles in the specific environment of the aquaculture road.The main work of this paper includes the following aspects:(1)In view of the insufficiency of the data set of obstacles in the aquaculture road independently collected in this paper,image processing technology is used to augment the data of the image,mainly using methods such as image smoothing,geometric transformation of the image,and color transformation of the image.Analyze the characteristics of the data set and general processing methods for these characteristics.(2)An improved data augment method is adopted at the input of the YOLOv4 algorithm,Mixup and Mosaic data augmentation are fused,which is recorded as Mosaic&Mixup data augment.After experimental comparison,the method has improved in terms of average accuracy and recall rate.(3)Improve the convolution in the backbone and replace the ordinary convolution with a depthwise separable convolution method.It is proved theoretically and experimentally that compared with ordinary convolution,depthwise separable convolution can greatly reduce the number of parameters generated by the algorithm model,thus improving the recognition speed.(4)In the neck structure,in order to balance the recognition speed and accuracy of the algorithm,the original PAN structure is replaced with the FPN structure,which reduces the calculation parameters in the network model again,and the calculation amount is also greatly reduced,and the recognition speed is further improved.(5)The Dilated Convolution is introduced at the prediction part to effectively improve the recognition capability of different target scales.After experimental comparison,the average accuracy and recall rate of recognition can also be improved.Finally,this paper fuses each improvement point to obtain the most effective algorithm model,whose recognition accuracy can reach 95.95%;FPS can reach 38,which can meet the needs of real-time detection;the recall rate reaches 94.08%,which proves that the algorithm satisfies the requirements for accurate and fast identification of obstacles on aquaculture roads,and the memory of the algorithm model is reduced from the original 245.9M to 33.4M,which makes it easier to implement the algorithm deployment.The implementation of this algorithm is conducive to promoting the application of self-driving feeding vehicles in aquaculture environments.By embedding this algorithm,obstacles on the road can be quickly identified,and the development of the aquaculture industry can be promoted.