Image-Based Sense and Avoid of Small Scale UAV Using Deep Learning Approach

Distance detection of target object is an important information for obstacle avoidance in many fields, such as autonomous car. When the distance of the obstacle is calculated, one can determine the potential risk of collision. In this paper, a monocular camera was utilized to get the distance from an incoming unmanned aerial vehicle (UAV) using deep learning approach. The distance detection of an UAV using You Only Look Once (YOLO) object detector was proposed in this study. The region which contain the detected UAV was processed into 100 by 100 pixel and was input into the proposed model to estimate the distance of the target object. For the proposed model, a Convolutional Neural Network (CNN) was adopted to solve the regression problem. First, the feature extraction based on VGG network was performed, and then its results was applied to the distance network to estimate distance. Finally, Kalman filter was used to improve the object tracking when YOLO detector is not able to detect UAV and to smooth the estimated distance. The proposed model was trained only by using synthetic images from animation software and was validated by using both synthetic and real flight videos.