Research On Visual Obstacle Avoidance Method Of Multi-rotor UAV Based On Time To Conllision

Multi-rotor drones are widely used in various scenes of human life,aerial photography,rescue,film shooting,circuit inspection,etc.Improving the ability of the multi-rotor drones to autonomous avoid obstacle is a key element for the safe flight.The environment perception and autonomous flight of drones have also been research hotspots in the industry.The focus of traditional obstacle avoidance research is to accurately obtain the distance from obstacles.In contrast,we model the rapid obstacle avoidance of animals and use a monocular camera to estimate the dense time to collision then we establish a local probability map to achieve the effect of UAV obstacle avoidance.The time to collision is used to comprehensively consider the distance information,the direction information and the motion information of the drone at the moment.It can uniformly deal with dynamic and static obstacles,providing a richer range for subsequent obstacle avoidance algorithms.We design and implement a dense TTC algorithm based on deep learning.The algorithm can output the TTC of each pixel.This part mainly contains the following innovations: In the network structure,the convolution structure part,we introduce the correlation layer to match the high-level features and enhance the relevance of the network.In the deconvolution part,the output mode of the multilayer pyramid is designed,which is related to the loss function.By setting the weights of each layer,the accuracy of the final output of the network model is guaranteed,and the training process of the network is accelerated at the same time.In the network training,we use the simulation data to solve the problem of missing data sets.We design and implement a local probability map algorithm based on TTC.This part contains two innovations,the fast query algorithm and the data structure that serves the query algorithm.The query algorithm can quickly query the relevant measurement frames containing various points in the current perspective and the corresponding uncertainty in the historical measurement frames,and then comprehensively evaluate the uncertainty of each point in the current frame to determine the final passable area.The data structure is a double linked list linear structure composed of pairs of edges and vertices,where edges represent the transformation relationship and uncertainty propagation relationship between the corresponding two vertices,and at the same time,the corresponding edges and vertices can be easily inserted and deleted.We design and implement a simulation platform for testing the autonomous navigation of drones based on Unreal engine.The platform can simulate the motion of the drone and generate the sensor data in real time,and at the same time provide a complex virtual scene with high fidelity,which is convenient for data collection and algorithm verification.The simulation platform was used to test the output effect of the dense TTC network model by comparing with the true value in different visual scenes.Subsequently,the simulation scene was used to verify the obstacle avoidance effect of the local probability map,and finally achieved a better result.