Research On The Vision System Of Tomato Harvesting Robots Working At Night

Harvesting robot is an effective way to realize automatic picking of fruits and vegetables,and it is a hot research point in researches on agricultural robots.Implementation of harvesting robot work at night,not only can it extend working hours of harvesting robots,but also improve harvesting efficiency.Recognition and localization are the main goals of fruits and vegetables harvesting robots.To realize automatic identification of tomato at night,this thesis focuses on the design method for the night lighting systems of tomato harvesting robots,image segmentation and recognition algorithms for tomato images captured in nighttime.The main contents and conclusions are as follows:(1)Design of night lighting system for tomato harvesting robot.Light source type,light source layout and image acquisition distance were used as the test factors,orthogonal table L18(6×36)was also used to arrange experiments.And an image segmentation evaluation index,which could determine the optimal night lighting system,was proposed.It was calculated as the ratio of F between intra-class and inter-class variances of foreground and background of tomato images captured at night.Testing results showed that,light source type and light source layout were two significant factors for night lighting system,while the image acquisition distance was not.Meanwhile,the F value of fluorescent lamp was 2.159,and that of the diagonal layout was 2.234.Therefore,in the current trial,the optimal light is fluorescent lamp,and the best light source layout is diagonal layout for the night lighting system of tomato harvesting robot.F of tomato images captured at night were compared with image segmentation results produced from image segmentation for same images using the OTSU automatic thresholding algorithm based on normalized R-G color difference.Comparative results showed that image segmentation results were better for images with higher F.The image segmentation evaluation index based on F and the design method for night lighting system are effective.(2)Preprocessing of tomato images at night.Image preprocessing is an essential step in the recognition of tomatoes,and the main effect is to reduce image noise and select color feature.According to the characteristics of tomato images in nighttime,a variety of noise reduction methods are analyzed.The results showed that Wiener filter can effectively reduce the image noise and preserve the edge characteristics of tomato images.In addition,the characteristics of each component of the tomato images in the color space were analyzed respectively.The results showed that the R-G color difference component in the RGB color space was used as the color feature of the tomato images in nighttime,and it was easy to divide and identify the tomatoes.(3)Design of image segmentation algorithms for tomato images at night.This thesis introduced the principle and implementation steps of OTSU automatic threshold method,K-means clustering method and Support Vector Machine.And the error rates and run time were used to evaluate the segmentation effect,the results showed that the OTSU automatic threshold method was used as the image segmentation method of tomato images in the nighttime environment,can be quickly and accurately divide the mature tomatoes.(4)Design of recognition algorithm for tomato images at night.In this thesis,we studied the recognition algorithms of single and double tomatoes at night,and focused on the recognition algorithm of overlapping tomatoes.The OTSU method based on red and green color difference was used to divide the mature tomatoes.And edge detection was realized by using Canny operator.Detected edges were divided into three kinds: outline edges,edges of lighting regions and overlapped edges.Thirdly,the nearest distance method was used to judge the anteroposterior position relationship of overlapped tomatoes.Finally,the front tomato without occlusion in overlapped tomatoes was recognized based on circle fitting.The results showed that: the accuracy of anteroposterior position relationship judgment was 96%,and the recognition precision for the front tomato without occlusion in overlapped tomatoes was 91.7%.In addition,the method can be used to determine the position of the overlapping tomatoes before the occlusion rate is less than 50%,and the tomato without blocked fruit can be directly identified,and the accuracy of anteroposterior position relationship judgment was 90%,the recognition precision for the front tomato without occlusion in overlapped tomatoes was 88.6%.The above work provided methods and technology foundation for the design of night lighting system and performance evaluation,tomato image segmentation algorithms and recognition algorithms at night of tomato harvesting robots,and it is helpful to improve the practicability of tomato harvesting robots.