Ground Objects Detection And Tracking For Low-power Embedded Devices On UAVs

UAV has an important strategic significance in the battlefield.Its main tasks are reconnaissance of military information,surveillance of targets,combat evaluation and other visual tasks.It also plays a key role in the automatic guidance of weapons.This paper mainly studies the detection and tracking of ground object on the UAV platform,and optimizes the algorithm design according to the characteristics of low-power embedded devices.In the high altitude vision of UAV,because of the wide sight distance,the distance between the interested object and UAV changes greatly.When the same object is at different distance,the object size is also different,and when the distance is relatively short,its rotation changes a lot.For the object detection task,it is difficult.In this paper,deformable convolution is used to solve this problem.When the object is far away,its size is far smaller than that of the close object.In the same picture,the object with large scale difference is more difficult for the object detection task.On the UAV platform,it can only carry low-power embedded devices.Due to the low power consumption of embedded devices,it is almost impossible to carry real-time object detection and tracking system.In fact,timeliness has always been a problem.In the rapid object detection algorithm designed in this paper,we mainly use residual structure technology,dilation convolution technology,depth separable convolution technology,deformable convolution technology,multi-layer feature fusion technology,etc.to design a rapid object detection network with good effect on ground object at high altitude.Its speed reaches 8.81 fps and average accuracy reaches 86.3% It has high practical significance to be able to deploy in the low-power embedded devices on the UAV.On the basis of object detection,the multi-object tracking in visual tasks is further studied.The color histogram and joint features of Io U of the object are extracted,the distance calculation matrix is completed,the motion prediction of the target is made by Kalman filter,and the feature matching of the detection target and the predicted object is completed by Hungarian algorithm.Combining the object detection algorithm and tracking algorithm,a low-power object detection and tracking system for UAV platform is designed,which can reach 12-15 fps in speed,and has certain practical value and significance.