Research On Key Elements Detection In Vision Based Factory Transportation AGV

With the increasing level of production automation and increasing labor costs,the AGV(Automatic Guided Vehicle)in the factory is gradually being widely demanded by all kinds of industries.At present,the commonly used AGVs mostly use electromagnetic navigation.However,due to the high cost of early construction,poor flexibility and equipment miscellaneous,it is difficult for AGVs to popularize in small and medium-sized enterprises.The AGV based on visual navigation is still not mature enough in China,not so easy to be put into production and application.In order to develop an AGV for unmanned transportation in the factory,this paper studies the key element detection technology in factory environment,analyzes the visual characteristics of key elements in details,and proposes a series of vision-based factory element detection methods,which greatly improves the detection efficiency.This article focuses on three key elements in factory using AGV:navigation signs,personnel and warning lines on the ground.The three types of elements have different visual characteristics,and the corresponding detection algorithms are also facing different needs and challenges:(1)Navigation signs usually occupy a small part in the image,and the shape,color,and texture characteristics are similar,which is easy to be confused and difficult to guarantee speed an accuracy at the same time.(2)Due to the distances and movements,personnel in factory have a large scale range,especially the shape and color feature.Also,the recall rate is very low when detecting personnel under cover.(3)The warning line has obvious and stable characteristics,so the detection accuracy is easier to get,but its requirement for detection speed is higher to ensure the smooth tracking of the AGV,so the calculation should be minimized in the algorithm.To solve these problems,this article designed effective detection methods for each element:in the navigation sign detection framework,using depth-wise convolution in YOLO v3 deep learning detection framework and improve the scale in detection layer,improve the detection speed by 40%on the premise of ensuring accuracy.And then use a deeper CNN network for fine classification;in the factory personnel detection framework,the channel attention mechanism is introduced and the IOU mechanism is improved to balance the detection effect at all scales,Which improves the accuracy rate by 1.7%.In the warning line detection framework,a variety of traditional image processing methods are compared and improved on the basis of the Hough transform,which improves the detection efficiency.In summary,this paper proposes effective detection methods for three key elements in the AGV usage environment in the factory transportation.Through experiments on multiple data sets,the effectiveness and the robustness of the proposed method is verified.