Learning Based Sense-and-Avoid Of UAVs

In recent years,the application potential of unmanned aerial vehicle(UAV)systems in military and civil fields has attracted much attention.As a result,the security and autonomy of the system are also the necessary conditions for the UAV systems to perform complex tasks.However,the limited communication capacity between air and ground and the weak man-machine interaction bring great challenges to the autonomous and security performance of the UAVs.Especially when dealing with unexpected obstacles,UAVs need to rely on sensors and controllers to realize autonomous sensing and autonomous avoidance.This paper mainly uses visual sensor as the main sensing source for lightweight UAV,aiming at exploring the sense-and-avoid control technology and autonomous learning method under the condition of its own load capacity and limited computing power.The main contributions and contributions are as follows:(1)This paper constructed the autonomous reactive control framework for the UAV sense and obstacle avoidance system.The framework covered the sense and action state description in the process of obstacle avoidance.We also put forward the ways of constructing the needed key modules for the implementation of the framework,and integrated the learning mechanism to achieve the reactive behavior from perceptual information to decision-making.The deep learning is used to realize the representation and dimensionality reduction of perceptual states and the reinforcement learning is used to build relationships between perceived states and avoidance actions.Training and learning methods are used to improve the adaptability of drones in an unknown environment.(2)A salient target autonomous detection algorithm applied to the UAV's forward sensors was proposed.Inspired by the mechanism of human visual attention,this paper achieved the salient target and background image segmentation by suppressing redundant background amplitude spectrum.This algorithm mainly used the discrete cosine transform and inverse discrete cosine transform theory.To solve the difficult problem in the field of feature representation,this paper applied the convolutional neural networks to set up the using saliency detection algorithm structure.The saliency detection was defined into a classification problem,and was realized by using sliding window method.Salieny detection experiments are validated using public data.Frequency domain algorithm detection results are 9.5 times faster than the general algorithms about running time,and deep network detection results are 14.3%higher than the general algorithms about the F_?index.(3)A representation and autonomous depth distance estimation algorithm for complex ground environment were proposed by using the coding framework.We mainly use the traditional encoder and decoder network to propose a multilayer automatic coding network algorithm,and use convolution and deconvolution neural network to estimate the depth distance.The complexity of the ground environment poses a challenge to the autonomous avoidance of UAVs.This paper used encoder and decoder network to build a deep encoding network and realized the feature representation of the environment.But the image reconstruction results ignore many details in the image.In order to fit for the obstacle avoidance,this paper puted forward the multi-scaled depth distance estimation algorithm based on the deconvolutional networks.The depth of the image was chosen as a representation of the complex environment and the public data set is used to verify the effectiveness of the algorithm.(4)The mapped relationship between UAV visual perception and UAV avoidance action is constructed using the reinforcement learning framework.The neural network are used to realize value function estimation for reinforcement learning.For UAV obstacle avoidance application,combined with the relations between state and action,this paper mainly depends on the reinforcement learning and reward technique to realize the mapping from image to UAV discrete and continuous action.We mainly used the neural network and deep neural network to fit the decision functions in the framework,and use the paired database to train the mapping decision network.This paper mainly constructs three kinds of learning mapping network(deep Q learning,deep double Q learning and Actor-Critic architecture network)to realize the mapping from the images to the discrete and continuous actions of the UAV.The RBF neural network and depth network are mainly used to realize the decision function fitting in the network,and the paired database is used to realize the training of mapping decision networks.The different algorithms are verified using the semi-physical simulation environment respectively.