Application Of Computer Vision In Accurate Object Detection,Localization And Tracking Using Unmanned Aerial Vehicle(UAV) Platform And The System Development

Major breakthroughs in computer vision and deep learning have greatly increased the applications of vision technology in the fields like object recognition,localization,tracking and estimation.Unmanned aerial vehicles(UAV),are able to fly under various atmospheric and geological situations due to the agility.Clearly,smart vision system based on UAV platform has great advantage for activities like smart inspection,traffic monitoring,surveillance and wild animals’ census.Owing to the variety and complexity of flying environment,the altitudes of UAV and the lighting in the open-air environment have great influence on the image quality,which further influence the perception of object characteristics and poses a big challenge for smart object(target)detection and real-time tracking.In order to solve the above problems,the study here has proposed an approach to improve the image quality and realized the smart multi-objects detection and real-time tracking with the application of deep learning using the main stream neural network architecture.Additionally,based on the researches on the image enhancement,object detection and tracking,a smart vision system integrated with the UAV platform has been developed and tested to verify its accuracy and reliability for the set scenarios.The mean works include:The image distortion caused by hardware like camera can be rectified through calibration of camera parameters using detected markers.Aiming at improving the digital information quality of the image which is related to the natural lighting factors and the information collection and transmission quality affected by the UAV altitudes and other environmental factors,the image information has been mathematically operated to reduce or eliminate so called “noise” by converting raw image into grey scale and by filtering the “noise” information.Specially,using image augmentation algorithm with self-adaption feature,a neural network architecture,which includes the sub network architecture for the considerations of the influence of lighting factors,has been proposed to improve the efficiency and accuracy of object detection.Three object detection algorithms(models)are chosen after the consideration of the characteristics of main stream detection algorithms(models)and comparisons of their detection efficiency.The training samples are from three datasets and then the images of the set scenarios are used for object detection.The most suitable YOLOv4 detection algorithm is selected after the analyses of the previous training and detection results.The anchors of the algorithms are re-clustered for another round of training and object detection to improve the detection accuracy.The methodology for detection accuracy improvement follows such a way that after the characteristics of layer information is analyzed and detection algorithm is improved after re-clustering,another round of training phases and detecting phase starts before the following detection accuracy analyses.In order to achieve real-time target tracking,the tracking by detection strategy is used to transform the "two-step" process of identifying and matching in target tracking into a "one-step" matching process.The target detection algorithm is responsible for the accuracy of tracking,and the target tracking algorithm is responsible for the matching degree of adjacent frames,which improves the generalization ability of the tracking algorithm.The network model constructed not only considers the commonly used tracking algorithms in the tracking algorithms.,Such as Kalman filter responsible for prediction,Mahalanobis distance and cosine distance responsible for calculating distance,Hungarian algorithm responsible for data association;also considered appearance feature information and combined with edge feature recognition Deepsort algorithm to test and compare YOLOv4 detection results and improved The target detection results ensure the reliability of the algorithm.In order to further improve the accuracy,the appearance feature extraction network in the Deepsort algorithm is retrained and tested.The architecture of the smart vision system integrated with UAV platform has been designed,developed and completed for real-time ground objects detection and tracking,and the accuracy and reliability of the system has been evaluated by both simulation and on-site tests.Taking into account the compatibility with the UAV flight control system,wireless transmission methods have been used for information exchange and transmission.Algorithms proposed in this study for image augmentation,object detection and tracking have been used for function realization.The system GUI is developed using PYQT5,and functions for real-time objects tracking and trace display have been realized by embedding Baidu module into the system.