Design And Development Of Agricultural Autonomous Navigation Rotary Tillage Robot Based On Deep Learning

In recent years,with the national emphasis on agricultural modernization,China’s agricultural development has made a great breakthrough,which has also driven the development of the agricultural planting industry in Ningxia.Through the research of the agricultural planting base in Helan County,Ningxia,the crop planting and harvesting process is manually operated by backward agricultural machinery and equipment and traditional human labor tools,which have not been updated for a long time,resulting in their backward technology,and traditional human labor tools are labor intensive and low efficiency.The design of intelligent agricultural robots can solve these problems to a certain extent.This paper designs and develops an autonomous navigating rotational tillage robot for an agricultural planting base in Helan County,Ningxia,and the details of the research are as follows:(1)The hardware structure of the rototilling robot is designed.Firstly,the chassis was selected,and the crawler mechanical tractor was chosen as the chassis of the rototilling robot,and the kinematic principle of the crawler chassis was briefly analyzed,taking into account the on-site construction environment of the agricultural planting base.The rotary tiller was studied,and the horizontal rotary tiller was determined as the rotary tillage system of the rotary tillage robot studied in this paper.The overall hardware part design of the rotary tillage robot chassis and rotary tillage system,camera,upper computer,lower computer and drive module provides the hardware basis for the rotary tillage robot to perform autonomous navigation.(2)The research and design of route navigation algorithm for autonomous navigation of rotary tillage robots is carried out.In this paper,a deep convolutional neural network based on residual structure is selected as the framework of the route navigation algorithm,and by designing the network structure of ResNet-34,the LeakyReLU function is determined as the activation function between layers,the cross-entropy loss function is selected as the loss function of the network,the L2 regularization is proposed to alleviate the network of the route navigation algorithm for rototillers The SGDM optimizer is selected to optimize the network,and the exponential decay learning rate method is selected to configure the optimal learning rate for the network.The image data collected from the first view of the rotating robot camera is used as the input of the lower computer,and the final output is changed into three nodes representing the straight,left and right driving actions of the rotating robot after processing by the improved ResNet-34 network model inside the lower computer,and the output prediction information is used as the input of the driving module to control the driving status of the rotating robot.(3)Research and design of obstacle detection algorithms for autonomous navigation of rotor-tilling robots.In this paper,the YOLO algorithm in regression-based target detection algorithm is selected to detect obstacles in the autonomous navigation process of the rototilling robot,and an improvement scheme is proposed to the original YOLOv5s network neck network to realize feature enhancement of non-neighbor layer information exchange and add SENet attention mechanism,which improves the recognition accuracy of the network model to detect small targets at a long distance as well as obscured targets.The image data collected from the first view of the rotating robot camera is used as the input of the lower computer to control the rotating robot to stop when an obstacle is detected.The experimental results of the autonomous navigation algorithm of the rotary tillage robot at an experimental site on campus are: the maximum lateral deviation of 11.6cm with the straight attitude as the starting attitude,the adjusted system error of 3.3cm by reducing the steering speed of the rotary tillage robot,the average lateral deviation of 2.05 cm,the standard deviation of 6.358 cm and the variance of 40.423.pedestrian,bicycle and car are selected as obstacles.The experimental results show that the improved network has significantly improved the level of obstacle detection,and the control of the rotary tillage robot stops working immediately when an obstacle is detected.The experimental results of the autonomous navigation and tillage system of the tillage robot show that the deep learningbased agricultural autonomous navigation tillage robot designed in this paper basically meets the design requirements.The experimental results of the autonomous navigation system and the rototilling system of the rototilling robot show that the deep learning-based agricultural autonomous navigation rototilling robot designed in this paper basically meets the design requirements.