Research On Deep Learning Algorithm Of Object Detection For Intelligent Robot

Intelligent robot is one of important carrier of artificial intelligence and has widely used not only in the industrial field but also in every area of our daily life and work.The object detection technology is one of the skills necessary for an intelligent robot to fulfill its mission and is also a difficult problem to be solved.Aiming at this issue,the object detection methods applied to intelligent robots are studied.The following is a list of the contents of the study:1.The traditional object detection algorithm is introduced-algorithm using sliding window followed by template matching.Because of this algorithm has many deficiencies,and deep learning algorithm showed its advantages in many image recognition and detection tasks,such as not dependent on human experience,better robustness and faster.Therefore,an object detection algorithm based on deep learning was proposed,in order to achieve better accuracy and real-time performance.2.Deep learning,especially convolutional neural networks,has great advantages in the field of image due to its unique structure.Object detection algorithms that based on convolutional neural network are also increasingly highlighting the potential.3.Based on Faster RCNN,a target detection algorithm suitable for this subject is designed.Firstly,a feature extraction network structure was constructed,which is composed of multiple convolutional layers,which are used to extract the original feature descriptions of input data.This process of feature extraction is completely automatic and requires no human intervention any longer.Furthermore,the convolutional feature contains more abundant information than the traditional features.Then,a region extraction network was constructed,which is also a convolutional network for receiving the original features as input,extracting the regions-of-interest of the image.It greatly cut down the time expense of selecting the proposal area;Then,a classification with location regression network was constructed which consists of an ROI pooling layer and full connectivity layers.The convolutional feature and the interest region of the image are input simultaneously.The two fully connected layers output the classification result and the border of the object respectively location information.A dataset of total of nine classes including both small objects and large objects is established.The pre-trained model and the self-built dataset are used to train the model.And then experiments are conducted.In most cases,the detection effect is very good,but there is still room for improvement.4.The detection of small objects has been a difficult problem and positioning accuracy needs to be further improved.Aiming at this issue,a multi-layer feature fusion method is proposed,and more detailed information effectively improves the detection ability of small objects.The object detection task of our subject includes not only the smaller objects to be captured but also the obstacles and reference objects with large size span.Taking into account the features extracted by Inception structure include a variety of scales,it was added to the algorithm;Using C.ReLU function as the activation function of the convolutional layers,the number of convolution kernels halved,so that the amount of computation was greatly reduced and speed up the detection speed ensuring the high accuracy.5.The platform of intelligent robot is introduced and the object detection platform based on Nvidia Tegra X2 embedded development board is established.The intelligent robot object detection system is designed and an interactive and easy-to-use software application is written.The test experiments are completed using intelligent robot target detection application.