Bio-Inspired Topological Navigation For Ground And Aerial Unmanned Vehicles

This thesis aims at developing bio-inspired topological navigation approaches to solve the self-navigation challenges faced by aerial and ground unmanned vehicles.The main work and contributions are as follow:1)In order to improve unmanned vehicles' place recognition ability,we propose to use convolutional networks to learn robust visual feature representations.Firstly,a novel multiscale feature pyramid is built with multiple pooling operations.Based on the feature pyramid,we then construct two network structures to learn visual feature descriptors.The first network structure fuses the multiscale features in the pyramid to form new visual descriptions,which have multiscale awareness.While,the second network structure learns an attention model from the feature pyramid,which can assign different weights to different local features.Finally,we trained the proposed networks in an end-to-end manner and compared with other networks.The results demonstrate that,the feature descriptors learnt from the two proposed network structures show improvements over other existing networks.2)Existing sequence based visual place recognition requires a storage-intensive image database for robust localization,while more storage-efficient odometry-based place recognition approaches can require a long travel distance to obtain an accurate localization.In order to obtain a fast and robust localization system with efficient map size,we present a novel particle filter-based localization system that adapts to varying degrees of map image densities,road layout ambiguity and visual appearance change.Firstly,the roadmap of the navigation environment is represented as a graph.Then,the proposed particle filter combines the geometric consistency between odometry and roadmap,and the visual similarity between the observed scene and the known scenes,to determine the vehicle's position.To improve the robustness of the particle filter under severe visual appearance variations,we also develop a vision reliability estimation algorithm to estimate the contribution of visual similarity to particle weights.We conduct comprehensive experiments and the results show that the proposed algorithm obtained high localization precision with short travel distance.We also introduced the path planning method to find a sequence of graph nodes to build the way to the destination.3)Inspired by the form of pigeons' cognitive map,a novel approach is developed to construct topological maps for aerial vehicle navigation applications.Firstly,we propose to build the topological nodes based on the saliency of visual landmarks.In this way,subregions with different level of saliency is grouped in different topological nodes.Then,we proposed to build the edges in the topological map to represent the reachability under compass guidance.Based on the topological map,a route planning algorithm is developed,which makes use of saliency nodes to construct routes to the destinations.As the planned routes make use of saliency nodes,they can guide the vehicle to the destination more reliably.The proposed route planning approach is also applicable in the navigation tasks that contain multi-destination and risk regions.4)An active bio-inspired navigation approach is proposed for aerial unmanned vehicles.Firstly,we adopted a monocular odometry with compass constraints to guide the vehicle across two connected nodes in the planned route.Then,we developed a sequence-based place recognition approach,which uses a Bayes filter to accumulate sequential belief in determining the position.Once the vehicle recognizes a place,a point feature match algorithm is performed to improve the localization accuracy.The localization result is then used to fix the odometry drifts as well as to adjust the orientation targeting to the next node region.Both simulated flight and real unmanned aerial vehicle flight experiments have been used to demonstrate the effectiveness of the proposed approaches.