Vision-based Target Detecting And Mapping System For UAVs

The application of artificial intelligence(Al)technology on carriers such as unmanned aerial vehicle(UAV)is becoming more and more popular.That rescue and reconnaissance UAVs perform target detection,location,search and rescue tasks in various challenging environments is a typical example of the combination of AI and UAV.In addition.UAV requires more reliable navigation and map construction techniques to cope with increasingly complex flight environments and operational tasks.This paper does research work on AI application,navigation and positioning and dense mapping of UAV.The main work and contribution in this paper is as follows:(1)Due to the problem that convolutional neural networks(CNN)with large parameters and high computational overhead cannot be applied to embedded devices in real time,this paper proposes a new CNN model compression algorithm based on filter pruning.By calculating the standard deviation of the filter in the convolutional layer to measure its importance,and pruning the filter and its corresponding feature map with less influence on the accuracy of the neural network,the computational cost can be effectively reduced.The experimental results show that the algorithm can compress the target detection model YOLOv2 by more than 50%and improve its operation speed based on the basic accuracy.(2)The increasingly complex application scenarios require more robust navigation and positioning functions of UAV.This paper introduces an inertial measurement unit(IMU)in the visual odometer(VO)system,and proposes a matching filtering algorithm based on geometric space,which can eliminate more mismatches,improve the accuracy of feature matching.And the Graph-Cut RANSAC(GC-RANSAC)algorithm is introduced in the visual front end to improve the calculation accuracy of the carrier pose.Experiments show that the improved visual inertial odometer(VIO)system has better precision and stability.(3)Independent visual odometer systems can only construct sparse maps,which can not meet the needs of path planning and navigation obstacle avoidance of UAV.Based on this,this paper introduces the SkiMap map construction framework in the VIO system,and proposes a fusion method of VIO and SkiMap based on the pose key frame.The efficient large-scale stereo matching(ELAS)algorithm is used to calculate the depth map of the binocular camera.The keyframes are extracted by calculating the amount of pose change between adjacent frames,and only the key frames are used for dense mapping.This method reduces the computational load and storage capacity of the dense map,and improves the efficiency of mapping.