Drones and Computer Vision

With the goal of communication innovation and application services, this project uses artificial intelligence to develop “Intelligent ANGEL base stations,” which can quickly establish a 3D UAV communication network with long-term hovering capabilities and self-optimization capabilities based on the dynamic demands of ground users. Including the following two key technologies:
1. 3D+R+V AI Data Analysis
– High-precision crowd estimation
– Dynamic crowd tracking
– Differentiated taget recognition

2. 3D+R+Edge AI-UAV communication, control, and computing platform
– Rapid reconstruction of radio signal map
– Software defined radio long-distance and adaptive commincation platform
– Design of long-hovering UAVs
– Multi-UAV commmunication network management

Image and video compression using deep learning technology. Currently, it is still under development but we believe it has tremendous potential. If it outperforms the certain image/video compression standards, it could possibly become international video standard after five years. And it could use as a compression framework for internal network system usage in the near future.

We focus on the development of integration of AI and UVA in warehourse environment.

(1)Object Detection and Tracking
(2)Depth Completion
(3)Parking Lot Vacancy Detection System
(4)AI Video Compression and Image/Video Restoration
(5)Medical Image Analysis
(6)Image Style Transfer
(7)IoT Wearable Device for Localization and Guidance System
(8)Automatic HVAC control system
(9)3D Skeleton Detection and Tracking
(10)Head Pose Estimation System
(11)UAV-Density Estimation
(12)3D Simultaneous Localization and Mapping(SLAM)

An autonomous intelligent drone technology, in which three functional components – computer vision, path planning, decision-making and behavior control, are developed and integrated as a self-contained system. The end-product is a software system that equips a brain to a drone for completely automated flight control. A visualized human-computer interface module is also developed to provide ground control station for mission management.

The techniques we developed include:
1. the camera calibration based on geometry principle
2. the one-shot fisheye camera calibration system
3. the multi-intensity with long-distance illuminated IR surveillance system and objects detection method
4. the smartglasses-based egocentric ranging/pointing system
5. the vision-based antenna pointing and registration system
6. multi-camera pedestrian positioning system

Recent methods in monocular depth estimation either have an over-smoothing effect on object boundaries or carry large computational costs to estimate depth values. We present a fast, accurate, and lightweight convolutional neural network (CNN) to estimate monocular dense depth maps from a single image augmented with a sparse depth map and its mask. We conducted experiments on a variety of challenging datasets, where we achieve state-of-the-art performance in terms of accuracy, network model parameters, and speed.

We propose a vision-based localization method. It integrates Visual SLAM and model-based localization method to achieve real-time, dirft-free, global positioning. The technology has been applied to navigation system for the visually impaired and a drone autopilot system that enables drone fly automatically in mixed indoor and outdoor environments.