AI in Air Traffic Management - NVIDIAon-demand.gputechconf.com/gtc/2018/presentation/s... · Thank...
Transcript of AI in Air Traffic Management - NVIDIAon-demand.gputechconf.com/gtc/2018/presentation/s... · Thank...
AI in Air Traffic ManagementChristian Thurow, Head of R&D at Searidge
WWW.SEARIDGETECH.COM/AIMEE
2
What is Air Traffic Control?
Motivation 1/3
3
Work Increase
Source: International Civil Aviation Organization,Civil Aviation Statistics of the World and ICAO staff estimates.
• Annual Growth: 6-7% • both #passengers and #flights • 2016: 3.7b passengers worldwide
Motivation 2/3
4
Motivation 3/3
5
Our Goals
5
Our Goals
• reduce controller workload
5
Our Goals
• reduce controller workload• increase situational awareness
5
Our Goals
• reduce controller workload• increase situational awareness• declutter workspace
5
Our Goals
• reduce controller workload• increase situational awareness• declutter workspace• provide additional surveillance data source (added safety)
6
What does Searidge do?
6
What does Searidge do?
7
8
Challenges
8
Challenges
• Building the NN
8
Challenges
• Building the NN • Training
8
Challenges
• Building the NN • Training • Inferencing Speed
8
Challenges
• Building the NN • Training • Inferencing Speed• Safety & Acceptance
9
1. Challenge: building the NN
• company policy: c++
9
1. Challenge: building the NN
• company policy: c++• first tried caffe, stayed with it
9
1. Challenge: building the NN
• company policy: c++• first tried caffe, stayed with it• first try with VGG16
9
1. Challenge: building the NN
• company policy: c++• first tried caffe, stayed with it• first try with VGG16• now VGG19 with custom layers for tracking (37 total)
9
1. Challenge: building the NN
• company policy: c++• first tried caffe, stayed with it• first try with VGG16• now VGG19 with custom layers for tracking (37 total)• superior performance over previous algorithm
9
1. Challenge: building the NN
• company policy: c++• first tried caffe, stayed with it• first try with VGG16• now VGG19 with custom layers for tracking (37 total)• superior performance over previous algorithm• problems: small objects
9
1. Challenge: building the NN
2. Challenge: Training
10
2. Challenge: Training
• Broad vs. Random Training Initialization?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?• How many annotations needed per site?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?• How many annotations needed per site?• Same training set for all airports or specific?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?• How many annotations needed per site?• Same training set for all airports or specific?• How many neurons, layers?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?• How many annotations needed per site?• Same training set for all airports or specific?• How many neurons, layers?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?• How many annotations needed per site?• Same training set for all airports or specific?• How many neurons, layers?• How many Epochs?
10
2. Challenge: Training
• Broad vs. Random Training Initialization?• How many annotations needed per site?• Same training set for all airports or specific?• How many neurons, layers?• How many Epochs?
10
11
3. Challenge: inferencing speed
12
4. Challenge: Safety & Acceptance
• safety first in ATC
12
4. Challenge: Safety & Acceptance
• safety first in ATC• need to prove performance
12
4. Challenge: Safety & Acceptance
• safety first in ATC• need to prove performance• regulator decides if system may be used operationally
12
4. Challenge: Safety & Acceptance
• safety first in ATC• need to prove performance• regulator decides if system may be used operationally• we treat ANN as human, same tests as for ATController
12
4. Challenge: Safety & Acceptance
13
Example Images and Videos
• list a couple of sample sites and show actual video
14
Example Images and Videos
15
Example Images and Videos
16
16
17
17
18
19
Future Work
• Optimal Flight Level Prediction • Optimal Aircraft to Gate Assignment • AI Controller Assist • many potential new application in ATC
Thank you!
HEAD OFFICE
19 Camelot Drive Ottawa, Ontario K2G 5W6
PHONE 613 686 3988 TOLL FREE 1 866 799 1555
EMAIL [email protected]
Thank you for your time.
I’ll be happy to answer any questions you may have.
21
Challenge: Annotation
Plattform Screenshots