Research And Application Of Object Detection Technology In Robotic Vision Systems

In the robot vision system,object detection is a technology to judge the category of all the objects in the image,and predict their location information.In traditional methods,the complex features obtained by integrating multi-class algorithms are less flexible,and the detection performance is difficult to improve.Although the deep learning method has good performance,its model is complex and its detection time is too long.It cannot meet the real-time requirements of industry,thus it is difficult to directly apply to industrial environment.In this paper,by observing the characteristics of industrial sorting robots,it is proposed that objects under fixed vision have scale invariant characteristics.In the convolutional neural network model based on deep learning,Faster RCNN model is chosen as the algorithm prototype for the dual consideration of detection speed and detection effect.It is the first time that the deep learning method introduced into the sorting system of industrial robots.Aiming at specific application scenarios,it is transformed into a special algorithm to meet the real-time requirements of industry.This paper divides Faster RCNN into four modules: convolution backbone network,region proposal network,Fast RCNN detector and multi-task loss.It focuses on improving the candidate box generation mechanism of Region Proposal Network.By screening the scale of the Ground Truth Bounding Box in the open data set,an additional training set is obtained,and the fine adjustment of the pre-training model is completed by using it.The improved algorithm make advantage of the method of feature dimension reduction,uses fewer feature vectors to generate candidate proposals,to reduce the computational load in the generation stage of candidate proposals,thus speeding up the detection speed.At the same time,around the scale invariance feature of the object,the conditions which the object needs to meet are discussed,and the applicable scenarios of the improved method are analyzed.Through the research and analysis of the candidate box generation process,an adaptive anchor generation mechanism is proposed.It improves the anchor generation mechanism with fixed scale in the region proposal network.According to the coordinate information,the mean value of the Ground Truth Bounding Box is calculated to set the scale of the central anchor.This improvement strengthens the detection ability of the area extraction network in the fixed field of vision.Even though the model uses fewer dimension feature vectors,the adaptive anchor mechanism can ensure that its detection performance does not decrease significantly.Through experiments,the improved Faster RCNN method is verified in real-time performance and detection effect.The results show that the detection speed is obviously accelerated,which meets the real-time requirements of the industry,and the accuracy of the method is significantly improved compared with the traditional method.Moreover,the improved method can also be applied to various scenarios,which can enhance the environmental adaptability of the robot and enhance its intelligence and technical level.