Jinwang Wang received the Ph.D. degree in Signal Processing Lab (SPL) at Wuhan University (WHU), advised by Prof. Wen Yang.
My research is on Computer Vision and Deep Learning. Particularly I am interested in object detection, localization and recognition based on aerial images (UAVs or satellites).
Researcher for Building Footprint Extraction in Aerial Images
Researcher for Object Detection in Aerial Images
Remote Sensing, 2019.
This paper addresses the influence of ambiguity of regression-based OBB representation methods for oriented bounding box detection, and proposes a mask oriented bounding box represetation (Mask OBB).
Sensors, 2019.
This paper presents a vision-based autonomous UAM with a 3DoF robotic arm for rotational grasping.
Geo-spatial Information Science, 2018.
A multi-scale hierarchical representation, i.e. binary partition tree, for analyzing large-scale UAV images was proposed.
2018 21st International Conference on Information Fusion (FUSION), 2018.
UAV-BD is a bottle dataset under UAV perspective with rotation bounding box.
2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2017.
A multi-scale Best-Buddies Similarity (BBS) method for matching objects in the UAV imagery was presented.
International Micro Air Vehicles Conference and Flight Competition (IMAV), 2017
We designed a wilderness search and rescue (WiSAR) system based on DJI M100 Unmanned Aerial Vehicle (UAV) and a ground station to search and rescue the survivors in wild.
I designed a novel arbitrary-oriented object detection framework suitable for aerial images based on Mask R-CNN to address the challenge of object rotations in the aerial images. This framework can generate oriented bboxes and horizontal bboxes at the same time. It has comparable performance on the DOTA Task1 and Task2.
In this project, I collected about 995 images which include DJI robots to train the optimized SqueezeDet, reached 40FPS on NVIDIA Jetson TX2 and the detection accuracy achieved 95.2%, and designed the target relative and absolute position localization algorithm by RealSense D435, the positioning accuracy can reach 2cm. It won 6th place in the DJI RoboMaster AI Challenge.
Demo
In this project, we collected about 20,000 images to build a UAV perspective person dataset (UAV-PP) and adapted detection algorithms YOLOv2, SSD and Faster R-CNN for person detection and chose the SSD as the detection algorithm because it’s balance on accuracy and speed, the detection accuracy achieved 88.92%.
Demo Poster
In this project, we designed a robot to carry blocks from the starting point to the working area. It contain four main modules, i.e. mechanical module, global navigation module, local navigation module and the designed controlling logic. It won 6th place in the DJI RoboMaster Mobile Manipulation Challenge.
Demo
In this project, we designed a UAV detection and automatic tracking system, including designed the PTZ control algorithm, implemented real-time tracking algorithm KCF to track UAV with low-speed flight in any direction. And We collected about 3,000 UAV images of DJI Phantom 3 for training the detection model of HOG+SVM, the detection accuracy achieved 78.6%.
In this project, we developed a programmable DC electric metamaterial using MCP42010 and MSP430 which can achieve the DC electric invisibility cloak, the DC electric concentrator and the DC electric illusion.
