A Case Study of Object Recognition from Drone Videos

Abstract

To study a potential autonomous drone's object recognition and reaction, we created a convolutional neural network (CNN) and used it to detect and count the empty parking spots in a parking lot taken from drone video footage. We first trained the network through supervised learning with snapshots of individual parking spots, from a previous drone footage, to correctly classify the spots as empty or occupied. Then we store the model to be used for detection and labeling of objects in new drone videos such as empty vs. occupied spots, as well as cars moving in and out of spots. We invented a video object referencing (VOR) to estimate object dimensions. After many rounds of tuning, we eventually achieve close to a hundred percent of accuracy. We concluded that adjusting batch size and epoch number could improve object recognition. We hope this research will contribute to tuning CNN for object recognition from drone videos to help with eventual autonomous drones.