| With the development of deep learning in recent years,the precision of object detection and tracking has strengthened a lot,making relevant industries boom,such as intelligent unmanned vehicle and UAVs.Due to its good flexibility and mobility,intelligent UAV systems have been widely applied in military,business and monitoring.However,many difficulties are lay on the way to the satisfactory performance.For example,it takes long time to complete the object detection process once by using convolutional neural networks,which is far from meeting the need for real-time detection.On the other hand,GPUs are require for running the networks,which is hard to get in UAVs.Meanwhile,it seems impossible to maintain stable tracking for the long time because of the deformation,illumination variation,fast motion,background clutter,rotation,occlusion,scale variation.In consideration of the needs and difficulties in various application conditions,a state-of-the-art object detection and tracking system has been designed to finish theses high-demanding missions with the help of UAVs' limited computing resources.The specific work of this paper is as follows:First,this paper designs an object detection network by imitate classic networks.The detection will be done in different scales and aspect ratio anchors.In pursuit of ever increasing performance,the parameters are adjusted.Second,the research about discriminative correlation filter for tracking is made.And the establishment of the object model has been improved.The results have implied that the improvement is effective,making the tracking more accurate.And the target's location information fusion algorithm is designed.The reasonability is tested.Finally,the design and test for the whole system are carried out.It is built on the ROS and are test on many test sets.By testing in the embedded system in UAVs,the result shows that it can runs in real time in different occasions and has fantastic performance,which means it can be put into practice. |
PDF Full Text Request