Visual Navigation Algorithm Based On Deep Learning

Deep learning provides a powerful visual perception tool for robotics applications.Compared with traditional pattern recognition algorithms on image application,the deep-learning-based visual perception algorithm discards the artificial feature engineering,adopts a data-driven scheme for adaptive feature learning.Deep learning models also have powerful nonlinear function fitting ability to achieve promising result in many different tasks.The data-driven characteristic of deep learning powers the algorithm with a high degree of self-adaptation and ability of evolution,making it suitable for robotic applications which in many cases with unstructured environments.In a universal vision-based mobile navigation application,the objectives of traditional vision algorithms are objection detection and localization,and map construction.Target detection and localization provides the mobile robot agent with the relative position of key objects or key points in the working environment as the motion tracking target;Map construction is global perception of the working environment,the special case of which is the construction of obstacle avoidance map.Map construction is essential for subsequent path planning and motion decision making.The target detection and tracking algorithm based on deep learning has the privilege on dealing with the low-level perception shortcomings of traditional algorithms,easily solving the problems of unstructured environment,occlusion and complex targets on an object detection task.Semantic segmentation based on deep learning classifies all pixels in semantic level,yielding a dense semantic map which can be used as simple as an obstacle avoidance map or an information enriched map for complex path planning application.The tradeoff of the powers fitting ability of deep learning model is a huge number of parameters.There are two main issues in engineering implementation and application.First,because of the deep structure,the model is large with a total parameter as much as millions,resulting a high latency and high computing resources requirement.Secondly,the huge parameter quantity of deep learning model requires a relatively same level amount of data to train the model to prevent serious Over-fitting phenomenon.Mobile robots in the field of robot applications have relatively poor onboard computing resources,and data collection for mobile navigation is difficult and costly.To tackle the problem of large parameter quantity and high delay of the model,it is necessary to optimize the model structure to reduce the model volume and overall calculation cost.For the difficulty of data source,Transfer Learning is adopted for model training,borrowing open source data for model pre-training and then collecting small volume of data for finetuning.The training scheme reduce the data volume while enhancing the generalization ability of the model.This paper mainly explores the research and application of visual algorithms for deep learning in mobile navigation.In response to the above problems,a multi-tasking network design is implemented to realize a unite neural network which meets multiple task requirements.The multitasking network design idea can increase the reuse rate of the convolutional neural network,reduce the model volume and computation latency.At the same time,the end-to-end neural network of visual-to-motion decision mapping based on deep learning is explored to realize look-and-move style visual navigation.For the training problem of the above design model,the idea of transfer learning is adopted.All models were tested offline and online,demonstrating the feasibility of deep learning in mobile robot visual navigation applications.Finally,in the summary and outlook section,further exploration of future work is provided.