Design On Object Identification Of Orchard Spary Robot Based On Deep Learning

The planting scale of orchard crops in China is expanding gradually.The traditional manual application and constant spraying machinery have problems such as high labor intensity,low efficiency,a large amount of pesticide waste and environmental pollution.Orchard spray robot is a new type of drug application machinery,which will develop in a more intelligent and accurate direction,among which the multi-classification identification of orchard target and non-target objects has become the research focus of spray robot.Recently,the rise of deep learning technology has opened up new ways for accurate detection and identification tasks.The classification identification and distance positioning of orchard spray targets through deep learning is realized in this thesis,and solves the problem of spray droplet interference to the detection during the spray operation.The main research contents are as follows:(1)The construction of the orchard spray robot perception system is completed,and designed the perception control program.The real-time information of the orchard spray operation scene is obtained through the binocular camera.The video flow data is processed by the embedded computer platform deployed in the model,the image is pre-foggy,and then the processed image is detected,identified and ranging,so as to obtain the accurate category and location information of the orchard target objects.(2)An improved YOLOv4 deep learning object detection model is proposed and the improved binocular visual ranging algorithm is combined to complete the identification and positioning of orchard targets.Based on the YOLOv4 original model,CSN of channel slice layer instead of pooling layer is used to reduce the loss of detail features during downsampling,and the path enhancement layer in feature fusion network is improved with BIPAN to reduce the leakage of small targets.The experimental results showed that the average accuracy and recall of the improved YOLOv4 target detection model were 0.61% and 0.68% higher than the original model,respectively.For the average distance positioning error within 10 m of the orchard target,it was 6.57% for fruit trees,1.70% for pedestrians,and 0.87% for telephone poles.The detection results prove that the improved YOLOv4 model combined with the binocular camera ranging algorithm can better realize the detection of the spray targets in the orchard environment,and provide a basis for the decision and control of the subsequent spray.(3)In order to eliminate the serious interference of water mist blur to the visual perception system,this paper proposes an improved AOD-Net deep learning defogging model based on the physical model of fog atmosphere as the preprocessing module of the YOLOv4 target detection model.The feature attention mechanism is adopted to improve the K value estimation module in the AOD-Net model,so that the network pays more attention to the areas with higher fog concentration in the image to achieve better defogging effect.Experimental data show that the improved AOD-Net model can effectively remove the interference caused by the image,improving the recognition accuracy and recall of orchard targets by 4.76 and 32.83 percentage points,respectively,while significantly reducing the distance positioning error of the binocular camera.In addition,the model also fully meets the requirements of real-time performance in performance.