Research On Semantic Segmentation Of Tennis Court Images

Semantic image segmentation of outdoor tennis scene is a key technology that needs to be solved for the development of intelligent sports service robots in the tennis court environment.It is also the algorithm basis for target recognition in a specific dynamic environment.Aiming at the limitation that existing ball-picking robots can only detect a single tennis goal,high-precision semantic segmentation research work is carried out on outdoor tennis court scene with high dynamic illumination,multi-target and multi-scale typical characteristics.The main research contents are as follows:1.Proposed a tennis court color segmentation model based on the HSV color gamut,and studied the effect of color threshold segmentation on outdoor tennis courts.Experiments were carried out to obtain the segmentation threshold and the performance and application scope of the color threshold segmentation method were analyzed.2.Established an image data set suitable for outdoor tennis court semantic segmentation tasks,including tennis,people,courts and other 9 types of targets.A tennis court semantic segmentation model is built on the Tensorflow deep learning framework.FCN algorithm was used to perform segmentation experiments on outdoor tennis scenes,and the accuracy of segmentation m IOU reached 49.47%.3.Aiming at the problem of FCN's inaccurate segmentation of small targets such as tennis balls and rackets,a series of convolutional neural network SACNet combined with Deeplabv3 algorithm is proposed.Loss,while using packet convolution to reduce the space-time complexity of SACNet.The test accuracy on the self-made tennis data set is 59.73%,compared with the existing Deeplabv3 model,the test accuracy is improved by 10.22%.4.A prospective target segmentation data set suitable for Mask R-CNN is established,and the joint matching model of Mask R-CNN and SACNet is used to perform instance segmentation on the three types of foreground targets: tennis,racket and human.Improved the accuracy of foreground target segmentation.The accuracy of the segmented m IOU on the test set reaches 89.88%.This paper studies the semantic segmentation of outdoor tennis court images,establishes a semantic segmentation dataset for outdoor tennis courts,proposes a SACNet model for multi-object segmentation,improves the segmentation accuracy for small targets and edges,and then uses Mask R-CNN model The segmentation accuracy of the foreground target is improved,which provides an effective method basis for the intelligent scene understanding and instance-level target recognition of the tennis court service robot.