My Ideas to Present
Created: 28 Mar 2023, 11:21 AM | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a")
Tags: knowledge, KnowledgeSharing
- Deep learning for Hyperspectral images, in earth observation satellites
- Relationship to GIS data, drone based GIS as part of future mobility - mapping of roads
https://www.smartnation.gov.sg/initiatives/strategic-national-projects/smart-urban-mobility
https://www.smartnation.gov.sg/initiatives/transport
Earth Observation / Remote Sensing
- GIS data
- Can be taken from many means - remote sensing != satellite, can include drone / other cameras
- Drones vs satellite imagery
- Drones are similar, but scale is abit different
- Drones vs satellite imagery
- What do EO images look like?
- Typically Hyperspectral images
- Nowadays also look at SAR images
Earth observation images
- Common use cases of such images include LULC, weather monitoring, disaster relief
- Deep learning and EO, 2 types:
- Processed on edge (in orbit, on drone)
- Benefits - save bandwidth … but I believe it is out of scope since it is highly unlikely that Continental will have its own fleet / constellation of satellites.
- Previously this was what I was working on using the nvidia agx xavier. Problem that I was solving was … show gif of results
- Post processed on ground
- Focus of today presentation
- Processed on edge (in orbit, on drone)
Benefits of satellites vs traditional methods
- Near real time (what is the time window?)
- Viable these days using satellite company’s API
Change detection
- Disaster relief
How all of this relates to AI, and to the automotive industry
-
Any specific papers?
-
HD road network estimation / mapping
- https://www.automotiveworld.com/articles/autonomous-vehicles-to-leverage-hd-maps-from-space/
- https://arlula.medium.com/satellite-imagery-powering-the-autonomous-vehicle-revolution-8c2e7cee5501
- https://www.nttdata.com/global/en/news/press-release/2019/april/build-high-definition-maps-for-autonomous-vehicles-from-space
- https://openaccess.thecvf.com/content/CVPR2022W/EarthVision/papers/Bahl_Single-Shot_End-to-End_Road_Graph_Extraction_CVPRW_2022_paper.pdf
- Single-Shot End-to-end Road Graph Extraction (CVPRW 2022)
- https://www.intelligence-airbusds.com/about-us/innovation/3d-mapping/ blimp based mapping system from airbus
- satellite imagery to tag road features in a bid to improve GPS navigation in all of these cases, as well as planning and providing disaster relief when road conditions may be dramatically altered.
From <https://www.tu-auto.com/on-the-road-with-future-in-vehicle-sat-navs/>
- Automated road information acquisition from remotely sensed data is always an interesting research area due to its promising values in various applications, e.g., autonomous driving (Wei et al., 2020,Yang et al., 2020), road network mapping making (Senthilnath et al., 2020), road network planning (Wang et al., 2021c,Zhang et al., 2020), traffic control and management (Zhou et al., 2020), map navigation and smart city construction (Chen et al., 2021b,Tan et al., 2020), etc.
From <https://www.sciencedirect.com/science/article/pii/S1569843222000358>
-
Motion prediction, Occupancy and Flow prediction challenges
-
Real time transportation updates?
- movement of goods and people, allowing transportation companies to optimize routes and schedules. This can help improve efficiency and reduce costs, ultimately leading to improved profitability.
- Pose estimation based on satellite image
- Geolocalization using ground-to-satellite cross-view image retrieval
- Knowing the exact location of a vehicle is critical for autonomous cars, and currently GPS systems are being used which the researchers claim suffer from limited precision and are sensitive to multipath effects — such as in “urban canyons” formed by tall buildings.
From <https://developer.nvidia.com/blog/satellite-images-help-track-a-vehicle/>
- The method consists of a Siamese network that learns feature embeddings suitable to matching groundlevel imagery with their corresponding satellite view
<!-- -->-
CVPR 2021
-
ECCV 2022
-
CVPR 2022
- Beyond Cross-view Image Retrieval: Highly Accurate Vehicle Localization Using Satellite Image [Yujiao Shi]
- Although this work was motivated by the poor accuracy of conventional image retrieval-based localization, we do not intend to replace the image retrieval-based localization technique. Instead, city-scale place retrieval can provide an initial estimate for a query camera. Our method then refines this pose estimate to higher accuracy
- TransGeo: Transformer Is All You Need for Cross-view Image Geo-localization [Sijie Zhu]
- Rethinking Visual Geo-localization for Large-Scale Applications
- Deep Visual Geo-localization Benchmark
- Beyond Cross-view Image Retrieval: Highly Accurate Vehicle Localization Using Satellite Image [Yujiao Shi]
-
WACV 2023
Possible business use case -
- Infrastructure Planning, Change Detection for Urban Mobility
- Mapping / Pose estimation for Autonomous Driving
Satellite based AI and how we can leverage on satellite images
Rainfall estimation, 60 rain monitoring stations, satelltie images
-
estimate the speed on the road
-
not finegrained speed detector - 1 to 5 bands on the road
-
try to predict the estimated arrival time based on certain distance
Possible presentation title
- “Satellites and Smart Cars: The Intersection of Earth Observation, AI, and Automotive”
- “Satellites and Smart Mobility: The Intersection of Earth Observation, AI, and Automotive”
- “From Space to the Road: The Power of xxxx”
- “From the Cosmos to the Car: xxx”
- “From the Cosmos to the Road: xxx”
From Outer Space to the Road - The Intersection of Earth Observation, AI, and Automotive
Satellites and Smart Mobility: The Intersection of Earth Observation, AI, and Automotive
TODO:
- Add lumelite 4 / TeLEOS-2 launch info, video snippet
- Add intro to my background as a satellite engineer (photos of me w the satellite), the projects that I am currently doing in Conti
Topic: Satellites and Smart Mobility - The Intersection of Earth Observation, AI, and Automotive
Speaker: Darius Tan
Speaker’s Biography:
Darius is an AI Engineer with a Master’s degree in Electrical Engineering focusing on Deep Learning and Computer Vision. Previously, he has spent over 3.5 years developing firmware for the control subsystem of 5 satellites in the National University of Singapore. Since joining Continental in Nov 2022, he has been working on Embedded AI for UWB applications and on Light-weighted Multitask Learning with the SGP AI and Data Engineering Team.