Detection, Recognition and Tracking Cars from UAV Based implementation of MobileNet- Single Shot Detection deep neural network on the embedded system By Using Remote Sensing Techniques

Tracking targets from the ground is difficult due to natural and artificial barriers, and in some cases, such as car detection, is dangerous, therefore, identifying targets using remote sensing is obvious. To achieve the purpose, the desired camera is installed on the unmanned aerial vehicle (UAV). with images processing on captured images from the camera, the system has used can identify the vehicle using aerial images and follow it if it is necessary. An important issue to this matter is the accuracy of the target detection. Therefore, efficient algorithms should be used in this field, and efforts have been made to use a deep neural network in this regard because it has the best performance rather than other methods. But using this network itself will cause other problems that are especially noticeable in realtime applications of the identification system. Because this type of neural network needs a lot of time to process information. Solving this problem will using strong hardware as much as possible, but these systems cannot be installed on the UAV due to their high weight and large power consumption. For this reason, in this paper, have tried to use pre-processing methods to identify possible moving targets and illuminate other parts of images to reduce the volume of data to make processing easier, and then the system can identify and track the car with the Light MobileNet-SSD network. This method is 25 times faster than other fast methods such as yolov3, and its loss rate is 0.02.