Research On Object Recognition And Location Technology Of Mobile Robot Based On Deep Learning

Object recognition and location technology is to identify the type of object in the input image or video and determine its position in the image or video,and then locate the object in space.The research core of target object recognition and positioning technology is target detection technology.The target object positioning based on mobile robots is to use the image position obtained by target detection technology,and then perform coordinate transformation to the space position of the object required by the corresponding mobile robot.In recent years,the research on target object recognition and location continues to heat up,and various application examples emerge endlessly,such as agricultural fruit and vegetable detection,pedestrian detection,uav remote sensing monitoring,etc.In this thesis,some typical target detection algorithms are evaluated comprehensively and YOLOv4 model is selected as the solution.At the same time in order to solve the YOLOv4 model under the condition of target dimension changes greatly the accuracy of problem,and the target detection of NMS algorithm under the condition of target overlap,shade,prone to error suppressed,leaving out the problem,research a kind of object recognition model based on YOLOv4 improvement goals and positioning method,this method enhances the precision of target detection,reduces the error suppression of the detection frame.The main research contents of this thesis include:1.Conduct a systematic research on deep learning algorithms,combined with the application conditions of the mobile robot itself,to determine the suitable deep learning target detection algorithm for this topic.In the model part of the target detection algorithm,in order to solve the detection accuracy problem of the YOLOv4 algorithm when the target scale changes greatly,this thesis introduces a new layer of SPP layer in the YOLOv4 model to increase the receptive field range.The scale of the pooling kernel has been modified to improve the information fusion ability of the model at multiple scales,reduce the information loss of the feedforward neural network in the network progressive process,and thus improve the detection accuracy of the model.Finally,the improved model was trained and tested on VOC public data set,and the improved model was used to carry out a real test on The VSTC mobile robot based on Jetson AGX Xavier computing platform to verify the feasibility of the improved model.2.In the non-maximum suppression(NMS)algorithm part of target detection,in order to solve the problem that the current NMS algorithm is prone to false suppression and missed detection when the target overlaps and occludes a lot,the information contained in this thesis according to the target detection box and real-time problem introduced the Manhattan distance of this investigation,put forward an improved algorithm of NMS.The improved algorithm was tested on the COCO data set,and finally on the physical platform of VSTC mobile robot to verify the validity and real-time performance of the algorithm.3.The improved method is tested on the real platform of VSTC mobile robot in this thesis.And according to the application scope of robot,simulating the fruit detection environment of agricultural robot,the corresponding algorithm experiment was carried out in the room,and the feasibility and real-time of the method was verified on the mobile robot platform.