Obstacle Avoidance By A Convolutional Neural Network In Consideration Of The Environment Recognition

Conolutional Neural Network was proposed in 1980th,It's a kind of locally connected Deep Architecture Neural Networks.It's special architecture consists of a set of learnable filters.In the ILSVRC 2012,it performed better than the traditional approach.Conolutional Neural Networks is widely used in the field of such as traffic sign recognition system in advanced driver assistance system(ADAS),weather prediction system based on image classification,facial point detection and pedestrain detecton.In the filed of Artificial Intelligence,obstacle avoidance based on convolutional neural network has successfully connect the input image from monocular camera to steering commands to a mobile robot.This method is easier than the traditional approach and has already made a success in obstacle avoidanse problem.However,this method is based on image classification,so when the distanec becomes longer,large amount of different images will be classifiled into the same cateogry.So in this article,we propose to divide the path into several pieces and train each of them by a certain CNN.The robot will do the environment recognition at the same time so that it will be able to choose which CNN to use according to the surrounding environment.In this article,we present an analysis of CNN and an experiment was performed using a crawler robot to check the accuracy of environment recognition by CNN.Also we introduced the cencept of boxel,to expand the filter from 2D to 3D,to mprove the architecture of Conolutional Neural Network.We introduced how to construct the weight matrix of this architecture.To insight the algorithm of CNN,we compared the traditional method in image processng such as rgb2gray,edge detector algorithm with CNN.And explained the reason why Maxout function is better than softmax function in the superficial layer of CNN.