Research On Obstacle Avoidance Strategy Of Agricultural Robotic Vehicles Based On Vision

The emergence of agricultural robotic vehicles based on human-machine cooperation,which reduces the labor intensity of workers and improves operating efficiency.Vision,as the most intuitive way to obtain environmental information by remotely operating smart vehicles and remote operators,is conducive to deepening the understanding of people and vehicles on the operating environment and promoting the human-machine cooperation process.It is of great significance to ensure the safety of operations and enhance the intelligence and automation of agricultural robotic vehicles.This paper proposes a vision-based approach to avoiding obstacles in agricultural vehicles.In the context of agricultural operations with complex backgrounds,vision is used to detect obstacles,locate them,and identify and classify obstacles.Qualitative concepts such as degrees are transformed into quantitative data;a collision prediction model in a dynamic environment is established.Based on the acquisition of obstruction information and the predicted collision results,a vision-based approach to obstacle avoidance strategies for agricultural robotic vehicles is realized.The following are the main contents of this article:1.Visual fusion-based obstacle detection.The operation images with obstacles are processed.Comparing Graph-based visual saliency(Gbvs)and Frequency-tuned salient(Ft)algorithms,Ft algorithm with good edge detection effect is used to generate saliency maps.The method of fusion of lidar and visual image information is described in detail.Lidar points and their corresponding pixel coordinates are clustered.Based on this,a restricted area growth method using Lidar points as seed points is proposed to implement obstacles segmentation in the image.The detection results show that the visual fusion detection method has better background suppression than the traditional segmentation method.2.Obstacle distance measurement.The visual distance measurement method of obstacles based on data regression modeling is described,and compared with the method of distance measurement combined with lidar and vision.Experiments show that within a distance of 30 meters,the maximum relative error of the distance measurement method based on data regression is 6.2%,the maximum relative error of the distance measurement method based on visual fusion is 2.1%,and the latter is more accurate than the former.3.Vision-based obstacle recognition.Combining the history of image recognition technology,the SVM algorithm and convolutional neural network algorithm(CNN)with high detection accuracy are selected to realize the identification and classification of pedestrians and agricultural vehicles with dangerous obstacles in the agricultural operating environment.This paper briefly describes the classification principle of SVM algorithm and elaborates the training process of CNN algorithm.A five-layer deep neural network structure was established.Training sample sets and detection sets were established.Based on the training results,the neural network training step was set to 20,the learning rate was 0.01,and the number of iterations was 320.The recognition results show that the detection accuracy of SVM is 89.6%,and the detection accuracy of CNN is 94.2%.4.Cloud-based obstacle avoidance strategy.According to the actual operation requirements of agricultural operation environment and agricultural vehicles,a dynamic environment obstacle avoidance strategy with speed control as the core is proposed.Explain the basic principles of the grid of obstacles in time and space,and establish a collision prediction model of a robotic vehicle in a dynamic environment based on the straight line prediction theory to determine the real-time collision position.Based on the cloud inference rules formulated based on expert experience and agricultural operating environment,the distance and risk degree of obstacles are taken as input,the speed control results are output,speed control strategies are established to realize speed control,and the A*algorithm based on dynamic guidance is selected.Implement path planning.Experiments show that the algorithm predicts and averages 0.17s,and the robotic vehicle speed control results are not affected by non-threat obstacles,and are in line with the speed cloud inference rules.The algorithm can realize real-time collision prediction and obstacle avoidance,has anti-interference ability and meets real-time requirements.