ATECT

Abstract

Object detection has been a very arduous task in the field of computer vision, artificial intelligence and machine learning for quite some time now. It has been an area of interest and study for many decades now. Object detection in aerial imagery has previously a bigger challenge due to lack of aerial imagery but now due to the advent of drone and aerial data, we can surmount this problem. For detection of vehicles and persons we need various angles and illumination conditions to capture a comprehensive environment.

Part 1: For vehicle and person detection in aerial imagery “ATECT” we propose the use of a very state of art and fast neural network architecture YOLO. The main challenge for our particular application was to bring object detection in real time to facilitate tasks such as that of surveillance. The other challenge was that the neural network models used to perform such complex detection are deep and require considerable computation power. So YOLO, though a little less accurate than another state of the art counterpart Faster R-CNN outperforms all other models when it comes to real time object detection.

Part 2: Our approach also involves the use of client server model for processing of the image stream sent by the UAV. As discussed above the models used to process images or perform detection are very deep and complex so we cannot do processing on mobile or on UAV. We need powerful GPUs to do object detection in real time.

So in short our FYP has to major components, the training of YOLO model for vehicle and person detection and the development of application and server side software for the communication of image stream. Then we would combine both for our final results.