Single Camera Robot Navigation Research Based On Convolutional Neural Network

At present,intelligent robots have been widely used in many fields such as exploration,rescue,material transportation,environmental management,and investigation.Constructing an intelligent robot road travel system that can achieve fast path navigation and obstacle avoidance is the core research content of intelligent robots one.At this stage,most intelligent robots,including smart cars,have used a combination of laser,radar,ultrasonic sensors,and visual cameras to implement obstacle detection and road following functions on road navigation.However,humans can accomplish obstacle avoidance and road following and other related functions only through their own vision systems,that is,our eyes,so intelligent robots should also be able to complete obstacle detection and road following using only visual sensor-cameras.Wait for work.Compared with the combination of multiple sensors,the use of a single camera to achieve navigation features economical,lightweight,low energy requirements and easy detection and many other advantages.This paper analyzes the current status of image classification methods and related applications based on convolutional neural networks,expounds the functions and meanings of each layer of neural network,and proposes a visual robot control algorithm based on endto-end deep convolutional network and designs a recursion and avoidance system.These systems validate the algorithm.The algorithm uses an end-to-end training pattern from the original map input image to the direction,achieving a high level of recognition.In order to verify the effectiveness of the algorithm and the corresponding system,this paper designs the two navigation systems for tracking and obstacle avoidance.The results show that the training error rate is 3.29% and the test error rate is 5.1% in the trail test.Among them,the training error rate was 1.8% and the test error rate was 5%,achieving very good image recognition effects.In the actual driving test,the robot can accurately follow the outdoor playground track line and indoor tin foil line,or accurately avoiding obstacles in the ground,which confirms the effectiveness of the proposed algorithm and the corresponding system.In addition,due to the traditional neural network image classification,there are some adjustment parameters that are difficult to identify or need to spend a lot of time to adjust the parameters to identify the image object,so consider the consistency of the image in the tracking and obstacle avoidance applications features,using a two-stream convolutional neural network,using the original picture and the optical flow picture respectively through the pre-trained network and then merging the scoring results,combining the two-stream convolutional neural network with the traditional image classification neural network to achieve the the accurate classification of the difficult-to-recognize images by the traditional neural network proves that this method can effectively improve the classification errors of traditional convolutional neural networks and improve the classification accuracy.