Research On Orchard Pedestrian Detection Method Based On Improved YOLOv3

Smart agriculture is the direction of future agricultural development.The development of agricultural machinery to smart agricultural machinery driven by artificial intelligence is a key link for the comprehensive transformation of traditional agriculture to smart agriculture.In recent years,unmanned agricultural machinery has helped to liberate the labor force and improve labor efficiency,and has achieved rapid development.In the safe operation of unmanned agricultural machinery,obstacle detection is one of the indispensable key technologies.Among them,pedestrian detection is the top priority of obstacle detection.Traditional target detection algorithms have low precision and recall rates,complicated models,and slow detection speeds.In recent years,deep learning algorithms have made considerable progress,and YOLOv3 algorithm is one of the most advanced algorithms in the field of pedestrian detection.Based on the YOLOv3 algorithm,this article conducts an in-depth study on how to achieve accurate and rapid detection of pedestrians in the orchard environment.The main research contents are as follows:(1)Elaborated the traditional target detection algorithm and deep learning target detection algorithm and principle.The YOLOv1,YOLOv2,and YOLOv3 algorithms for deep learning YOLO are introduced in detail,and the YOLOv3 algorithm is compared with other mainstream algorithms to clarify the superiority of YOLOv3 algorithm performance such as accuracy and speed.Theoretical basis.(2)According to the characteristics of the orchard environment,an improved YOLOv3 pedestrian obstacle detection algorithm is proposed.First of all,the YOLOv3 backbone network Darknet53 was replaced with ResNet50 network,ResNet series network is easier to expand according to the characteristics of different environments and scenarios,and ResNet50 network has more advantages in speed and accuracy.Then,introduce deformable convolution DCN v2 to replace the 3x3 original convolution of the stage5 part of ResNet50 network,which is beneficial to the balance of accuracy and speed.Finally,the regularization technology DropBlock module is added to the FPN part,which improves the model generalization ability.(3)Experiments verify the performance of the improved YOLOv3 pedestrian obstacle detection algorithm.Using the open data set of orchard pedestrian detection of the National Robotics Engineering Center of Carnegie Mellon University,based on the Tensorflow deep learning framework,the comparative experiments of the YOLOv3 algorithm before and after the improvement are carried out,mainly for small,medium and large targets And comparison experiments of pedestrians in different orchards without shelter,less shelter and more shelter.Through the analysis of experimental data,the average precision and recall of the improvedYOLOv3 orchard pedestrian detection model reached 96.21% and 91.12%,and the detection speed was increased by 3 times,and the detection speed was 52.3 frames per second,and the model parameter amount was reduced.Four times,this method has better real-time performance and generalization ability,and the model has better robustness,which can meet the real-time detection requirements of unmanned agricultural machinery pedestrians in orchards.