An Optical Flow-Based Complex Navigation Method Inspired By Insect Vision

Insects possess the capability to navigate precisely in a complex natural environment.The study of insect’s navigation behavior can provide novel strategies for the design ofefficient navigation systems. Currently, many physiological experiments have shown thatsome insects can use optical flow (OF) to implement various navigational tasks. Therefore,by deriving OF navigation strategies from insects, novel visual navigation methods maybe developed that would allow a robot equipped with such a system the same capabilityfor precise navigation as the insect.This thesis investigates the visual navigation strategies of insects according to someexperiments of the insect’s navigation. Specifically, the insects use OF to measure theirego-motion. The traveled distance and current position are then obtained by accumulatingmotion at each step through path integration. In order to correct errors from pathintegrating, the insects store a sequence of obviouse landmarks situated at specificdistances during the flight, and thus measure the position errors by using continuouslyvisual landmark matching. Especially, the insects can implement OF matching to obtainaccurate position when landmark is hardly distinguished.Learning from the insects’ OF navigation strategies, this thesis proposes abio-inspired complex navigation method based on OF. The complex navigation method iscomposed of an OF navigation method and an OF aided navigation method. The OFnavigation method is used to measure single-step motions along a path. The positioninformation is then obtained by path integration. To overcome cumulative errors, the OFaided navigation method is employed to assist the navigation system in estimating thesecumulative errors through OF matching. As a result, the proposed complex navigation canprovide accurate position information without interference from cumulative errors yetdoing so with low computational effort.The thesis first discuss the OF navigation method, which employs an OF methodinspired from insect vsion to measure its ego-motion. This OF method uses animage-interpolation algorithm to estimate single-step motions of camera with a simple andnon-iterative procedure. According to this OF method, we also develop a new OF methodby setting the entropy of image difference to its extremum. The proposed OF method ismore suitable to measure single-step and multidimensional motions of camera. In order to implement an aided navigation compatible with the above OF navigationmethod, a new OF-based kalman filter (KF) is presented to iteratively implement OFmatching. The aided navigation method using this KF can precisely estimated and thuscompensate for cumulative errors. In operations, the measured OF is derived from theactual input image, whilst the predicted OF and its slopes are computed from the predictedinput image. They are input into the update equations of the KF to generate the errorestimate. Moreover, by modifying the insect-inspired OF method used in the OFnavigation, a new OF with linear variation and rotation invariance is defined to generatethe measured and the predicted OF required by the OF-based KF. In addition, entropy isalso an important feature to measure image information. This thesis presents another aidednavigation method using the image entropy and the exterend KF.The proposed navigation methods were tested in the simulations by using real aerialimages with different noises and textures. Also, an actual mobile robot was used to movealong different trajectories in an indoor environment, and captured images from the ceilingto implement the complex navigation. Both the simulational and experimental results havedemonstrated the proposed complex navigation is efficient for guidance over longdistances and times.The navigation methods involve many inner-product operations. Therefore, wepresent a novel fast algorithm to compute inner products rapidly via transforming the innerproduct into a first-order moment. This algorithm is very efficient for the digitalconvolutions.