Grant: Alliance Grants
Université de Montréal
$35,500.00
Deep neural network uncertainty estimation for safe integration into autonomous driving autonomy systems
EF 2024-25
Résumé
Deep learning based perception systems, such as object detectors, have shown incredible improvements in
performance in recent years. It is desireable to incorporate these tools into robotic systems such as self-driving
cars. However, in order to do so requires that these models also have the ability to output reliable confidence
and uncertainty measures associated with their predictions. One particular challenge with deep neural networks
is that there are several sources of uncertainty that are difficult to disentangle. For example, an object might be
very occluded in which case it would be difficult for even a human to detect it, or the object might of a type
that is not seen in the training data. In both cases, the model will have high uncertainty about the about but for
very different reasons. Bayesian Neural Networks attempt to overcome this problem by outputing distributions
that approximate the optimal Bayesian estimate. The issue with most of these methods is twofold: 1) they do
not scale to the types of problems that are relevant in robotics, 2) they often provide uncertainty estimates that
are uncalibrated, meaning that they don't faithfully represent the actual uncertainty in the underlying process.
The objective of this project is threefold. First, we will build new algorithms for estimating sharp and
well-calibrated uncertainty estimates in deep learning based object detection. Second, we will demonstrate that
these uncertainty estimates can be used in other downstream components of a robotics autonomy stack, such as
sensor fusion and risk-based planning. Third, we will show that these methods are applicable to the more realistic continual learning setup where not all data is assumed to be available at once. We will demonstrate this on well-established datasets and also on a full-scale autonomous vehicle.
Programme
Alliance Grants
Bénéficiaire
Emplacement
Quebec, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
580895-2022
Name nom
Paull, Liam LD
Department département
Informatique et recherche opérationnelle
Organization id
63
Institution établissement
Université de Montréal
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2022
Award amount
35500
Program id
ALLRP
Program name en
Alliance Grants
Program name fr
Subventions Alliance
Group en
RESEARCH PARTNERSHIPS
Group fr
PARTENARIATS DE RECHERCHE
Committee code
96
Committee name en
RPP Internal Decision Cttee
Committee name fr
Cté de décision interne - PPR
Area of application code
700
Area of application group en
Transportation systems and services
Area of application group fr
Services et systèmes de transport
Area of application en
Transportation systems and services
Area of application fr
Services et systèmes de transport
Research subject code
2600
Research subject group en
Robotics
Research subject group fr
Robotique
Research subject en
Robotics
Research subject fr
Robotique
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Deep neural network uncertainty estimation for safe integration into autonomous driving autonomy systems
Keywords
autonomous vehicles; autonomy; deep learning; object detection; robotics; robot perception; uncertainty estimation
Application summary
Deep learning based perception systems, such as object detectors, have shown incredible improvements in
performance in recent years. It is desireable to incorporate these tools into robotic systems such as self-driving
cars. However, in order to do so requires that these models also have the ability to output reliable confidence
and uncertainty measures associated with their predictions. One particular challenge with deep neural networks
is that there are several sources of uncertainty that are difficult to disentangle. For example, an object might be
very occluded in which case it would be difficult for even a human to detect it, or the object might of a type
that is not seen in the training data. In both cases, the model will have high uncertainty about the about but for
very different reasons. Bayesian Neural Networks attempt to overcome this problem by outputing distributions
that approximate the optimal Bayesian estimate. The issue with most of these methods is twofold: 1) they do
not scale to the types of problems that are relevant in robotics, 2) they often provide uncertainty estimates that
are uncalibrated, meaning that they don't faithfully represent the actual uncertainty in the underlying process.
The objective of this project is threefold. First, we will build new algorithms for estimating sharp and
well-calibrated uncertainty estimates in deep learning based object detection. Second, we will demonstrate that
these uncertainty estimates can be used in other downstream components of a robotics autonomy stack, such as
sensor fusion and risk-based planning. Third, we will show that these methods are applicable to the more realistic continual learning setup where not all data is assumed to be available at once. We will demonstrate this on well-established datasets and also on a full-scale autonomous vehicle.