Human Centric Spatio Temporal Action Localization

Megvii(Face++) Team

Find Tiny Instance Segmentation for Autonomous Driving

Megvii

(Face++) Team

UAV to UAV Target Detection and Pose Estimation

The objective of this thesis is to investigate the feasibility of using computer vision to provide robust sensing capabilities suitable for the purpose of UAV to UAV detection and pose estimation using affordable CCD cameras and open coding libraries. We accomplish this by reviewing past literature about UAV detection and pose estimation and exploring comparison of multiple state-of-the-art algorithms. The thesis presents implementation studies of detection approaches including color-based detection and component-based detection. We also present studies of pose estimation methods including the PosIt algorithm, homography-based detection, and the EPFL non-iterative method. The thesis provides a preliminary strategy for detecting small UAVs and for estimating its six degree of freedom 6DOF pose from image sequences within the prescribed airspace. Discussion of its performance in processing synthetic data is highlighted for future applications using real-life data sets.

Fiducial Marker Detection and Pose Estimation From LIDAR Range Data

Light Detection and Ranging LIDAR systems are three dimensional 3D imaging sensors applied for mapping terrain, measuring structural dimensions, and navigating robots. Pulsed laser rangefinders provide precise range measurements that require an estimate of sensor pose for transformation into world coordinates. Pose information is frequently provided with extrinsic sources such as Global Positioning System GPS or an Inertial Measurement Unit IMU. Unreliable signal availability for GPS in military environments and the high cost of IMUs limit the employment of these extrinsic sources. Determining pose intrinsically by detecting landmarks in the environment within the sensor data is more ideal. Fiducial markers with known geometric dimensions and orientation provide a means of estimating LIDAR pose and registering data. Presented is a method for landmark detection and pose estimation within range data. Cylinder, cone, and sphere geometries are assessed for use as fiducial markers. The detection algorithm extracts geometric features from LIDAR point data and tests for fit to a fiducial marker model. Geometric feature extraction compresses the data set and leads to a potential intrinsic registration method using environment landmarks. The detection accuracy and pose estimation precision are examined with terrestrial LIDAR range data captured in various outdoor street environments.

Photorealistic Image Generation for Satellite Pose Estimation Using Generative Adversarial Networks

In autonomous satellite servicing operations, pose estimation is an integral process to guide the servicing satellite for rendezvous and capture of the satellite to be serviced. Convolutional Neural Network CNN-based methods show promise in satellite pose estimation. In order to train CNNs for pose estimation, sufficient quantity and quality of real training imagery that is labelled with detailed pose data are required. Such images are either unavailable or very costly to produce, often forcing augmentation using computer-generated or synthetic image datasets. In order to enable CNN-based pose estimators to fulfill their robust and efficient potential, one may draw from the distribution-matching ability of the Generative Adversarial Network GAN to modify an existing training dataset of synthetic imagery based on the characteristics of markedly fewer real images. This research focuses on the Cycle-Consistent GAN CycleGAN architecture for its strength in such style transfer tasks. Both a geometrically simple proof-of-concept object and the on-orbit images of a small satellite are employed for photorealistic image generation using CycleGAN and training of a simple CNN pose estimator. Resulting improvement to real image pose estimation accuracy of this CNN when trained on such photorealistic imagery vice synthetic imagery provides valuable insight to future applications of the implementation of CycleGAN for such training data generation.

Improving Automated Aerial Refueling Stereo Vision Pose Estimation Using A Shelled Reference Model

Automated Aerial Refueling of Unmanned Aerial Vehicles is vital to the United States Air Forces continued air superiority. This research presents a novel solution for computing a relative 6 degree-of-freedom pose between the refueling aircraft and a tanker. The approach relies on a real time 3D virtual simulation environment that models a realistic refueling scenario. Synthetic imagery is processed by computer vision algorithms that calculate the sensed relative-navigation position and orientation. Pose estimation accuracy and computational speed during registration improve though the use of a shelled reference model. The shelled model improves computational speed of pose estimation at the refueling position by 87 and accuracy by 36 when compared with a full reference model. To ensure proper simulation of computer vision concepts, this research quantifies the effect Multi-Sample Anti Aliasing implemented in the virtual stereo cameras on camera calibration and pose estimation. A combined shelled model and Multi-Sample Anti Aliased approach leads to position estimation errors less then 7cm and orientation estimation error less then 1 .

Object Detection with Deep Learning to Accelerate Pose Estimation for Automated Aerial Refueling

RPAs cannot currently refuel during flight because the latency between the pilot and the aircraft is too great to safely perform aerial refueling maneuvers. However, an AAR system removes this limitation by allowing the tanker to directly control the RP A. The tanker quickly finding the relative position and orientation pose of the approaching aircraft is the first step to create an AAR system. Previous work at AFIT demonstrates that stereo camera systems provide robust pose estimation capability. This thesis first extends that work by examining the effects of the cameras resolution on the quality of pose estimation. Next, it demonstrates a deep learning approach to accelerate the pose estimation process. The results show that this pose estimation process is precise and fast enough to safely perform AAR.

GENERAL PURPOSE PROBABILISTIC PROGRAMMING PLATFORM WITH EFFECTIVE STOCHASTIC INFERENCE

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

APRIL 2018

更多

FINAL TECHNICAL REPORT

APPROVED FOR PUBLIC RELEASE; DISTRIBUTION UNLIMITED

STINFO COPY

AIR FORCE RESEARCH LABORATORY

INFORMATION DIRECTORATE

收起