REAL-TIME VISUAL DETECTION AND TRACKING FRAMEWORK USING DEEP CONVOLUTIONAL NEURAL NETWORKS FOR MICRO-AERIAL VEHICLE

Abstract

Deep learning has achieved great success in computer vision tasks like object detection, tracking and segmentation. Attempts have been started to adopt deep learning algorithms for object detection and tracking on small UAVs. However, due to the enormous amount of processing power required in real time, there are still many challenges in applying deep learning algorithms on these platforms. In this thesis, I have implemented a Detection-Tracking framework with convolutional neural networks. The framework is implemented on an Intel UP Board processing unit and a UAV platform. Instead of processing images completely on-board, we handle tasks demanding much computational power in the off-board computer, while we handle light-weighted algorithms on-board. Images from the on-board camera are transmitted to an off-board Laptop via wifi network. The off-board laptop is equipped with an NVIDIA GTX 1060. The current framework is able to do detection and tracking autonomously in real time.