Object Detection And Autonomous Obstacle Avoidance Of Multi-rotor UAV

Using multi-rotor UAV for object detection task has great value in military reconnaissance,post disaster search and rescue,maritime rescue and other tasks.The object detection of multi-rotor UAV is also the basis of target tracking,autonomous obstacle avoidance and other tasks.It has important scientific significance and practical value for the object detection and autonomous obstacle avoidance of multi-rotor UAV.Due to the limited volume and load capacity of the multi-rotor UAV,the processor memory and computing capacity deployed on it are limited,which results in the limited speed when using the multi-rotor UAV for object detection task.In addition,aerial images have many problems,such as object occlusion,large scale change,large image field of view,small single object,complex background environment and so on,which lead to the low accuracy of object detection task using multi-rotor UAV.Traditional object detection algorithms,such as Viola-Jones detector,hog pedestrian detector and deformable component model,usually detect objects according to the steps of region selection,feature extraction and classification regression.The traditional model is relatively simple,it is difficult to extract the deep features,and the effect of object detection for multi-rotor UAV is poor.In recent years,the object detection algorithm based on deep learning: R-CNN series algorithm,SSD algorithm and YOLO series algorithm have obvious advantages over the traditional algorithm.Through the comprehensive analysis and comparison of these common algorithms,YOLOv3 algorithm has the comprehensive performance advantages of speed and accuracy,and is more suitable for deployment in mobile devices such as multi-rotor UAV.However,the original YOLOv3 algorithm still has a large amount of calculation.This paper mainly aims at the characteristics of multi-rotor UAV,improves the YOLOv3 algorithm,and proposes a more suitable YOLOv3-CPA object detection algorithm for deployment of multi-rotor UAV.In order to reduce the computation and memory consumption of YOLOv3,channel pruning is used to quantify the channel weight in YOLOv3,and the redundant channel is deleted.Experiments show that the size and parameters of the model can be reduced by 80%,the number of floating-point operations(FLOPs)can be reduced by 70%,and the forward inference speed is about twice as fast as the original model.In order to reduce the influence of complex background of aerial image on object detection of multi-rotor UAV,improve the mean Average Precision(mAP)of object detection,and make more effective use of computing resources,adding channel attention mechanism and spatial attention mechanism to YOLOv3,and give different weight to the features of aerial image,so that the network has more excellent feature selection ability.The network can pay more attention to important features and suppress non important features to improve the efficiency and accuracy of object detection.In order to solve the problem of object occlusion in aerial images,the non maximum suppression(NMS)strategy in YOLOv3 is changed to the combination of Softer-NMS and Soft-NMS,so that the problem of low recall rate caused by target occlusion in target detection is improved,and the position of boundary box is more accurate.The experimental results show that the recall rate of the model can be increased by 13% through the combination of Softer-NMS and Soft-NMS.By adding attention mechanism and NMS strategy change,the improved YOLOv3-CPA algorithm can improve the mAP by about 5% compared with the original YOLOv3 algorithm.In the stage of model deployment,TensorRT,a high-performance deep learning inference optimizer developed by NVIDIA,can improve the forward inference speed of YOLOv3-CPA algorithm by 2-3.4 times under different precision.When the final improved YOLOv3-CPA algorithm deploys NVIDIA TX2 and Xavier on multi-rotor UAV,the detection frame rates of 416 × 416 size pictures can reach 24.1FPS and 47.6FPS respectively,basically meeting the real-time requirements.In addition,the improved YOLOv3-CPA algorithm and deep convolution neural network are used to verify the application of autonomous obstacle avoidance of multirotor UAV.The image information is collected by the binocular camera of the multirotor UAV,and the expected attitude angle and throttle volume of the multi-rotor UAV are predicted by using the designed deep convolution neural network.In a specific scenario,the success rate of autonomous obstacle avoidance of the multi-rotor UAV is about 85%.The improved YOLOv3-CPA algorithm is used to detect specific types of obstacles.The multi-rotor UAV can detect visual obstacles,and the obstacle boundary box is used to optimize the path of autonomous obstacle avoidance.In specific scenarios,the success rate of autonomous obstacle avoidance of multi-rotor UAV is about 94%.The effectiveness of YOLOv3-CPA algorithm is further proved by autonomous obstacle avoidance.