Robust multimodal sensor fusion for autonomous driving vehicles
Techniques are disclosed for using neural network architectures to estimate predictive uncertainty measures, which quantify how much trust should be placed in the deep neural network (DNN) results. The techniques include measuring reliable uncertainty scores for a neural network, which are widely used in perception and decision-making tasks in automated driving. The uncertainty measurements are made with respect to both model uncertainty and data uncertainty, and may implement Bayesian neural networks or other types of neural networks.
1. An electronic control unit (ECU), comprising:
a plurality of a neural networks, each one of the plurality of neural networks having an input coupled to a respective autonomous vehicle (AV) sensor and configured to output distribution data representing a set of different classes used in accordance with an AV application environmental model;
a plurality of uncertainty estimation units, each one of the plurality of uncertainty estimation units being configured to calculate uncertainty estimate values comprising an aleatoric uncertainty value and an epistemic uncertainty value using the distribution data output from a respectively coupled one of the plurality of neural networks, the aleatoric uncertainty value being with respect to data provided by the AV sensor coupled to the respectively coupled one of the plurality of neural networks; and
a control unit configured to:
gate a contribution of data provided by each AV sensor to the distribution data output by each one of the plurality of neural networks in accordance with a gating function that utilizes respective uncertainty estimate values calculated for each one of the plurality of neural networks; and
generate the AV application environmental model using the gated distribution data.
2. The ECU of claim 1 , wherein each one of the plurality of a neural networks is a Bayesian deep neural network (B-DNN).
3. The ECU of claim 1 , wherein each one of the plurality of a neural networks is a non-Bayesian deep neural network (DNN).
4. The ECU of claim 1 , wherein each one of the plurality of uncertainty estimation units is configured to calculate respective uncertainty estimate values using predictive distributions obtained from multiple Monte Carlo forward passes through a respectively coupled one of the plurality of neural networks by sampling weights from a learned posterior distribution of each respectively coupled one of the plurality of neural networks.
5. The ECU of claim 1 , wherein the control unit is configured to monitor the uncertainty estimate values over time, and to flag an AV sensor as being faulty when a respective aleatoric uncertainty value associated with the AV sensor exceeds a threshold aleatoric uncertainty value.
6. The ECU of claim 1 , wherein the control unit is configured to monitor the uncertainty estimate values over time, and to initiate a re-training sequence for the plurality of neural networks when an epistemic uncertainty value associated with one or more of the plurality of neural networks exceeds a threshold epistemic uncertainty value.
7. An autonomous vehicle (AV) system, comprising:
a memory configured to store instructions; and
one or more processors configured to execute the instructions to:
receive sensor data via each one of a respectively coupled autonomous vehicle (AV) sensor and output sets of distribution data, each one of the sets of distribution data representing a set of different classes used in accordance with an AV application environmental model;
calculate uncertainty estimate values comprising an aleatoric uncertainty value and an epistemic uncertainty value for each one of the sets of distribution data;
gate a contribution of data provided by each AV sensor to each one of the sets of distribution data in accordance with a gating function that utilizes respectively calculated uncertainty estimate values for each one of a plurality of neural networks; and
generate the AV application environmental model using the gated distribution data.
8. The AV system of claim 7 , wherein the one or more processors are configured to output the sets of distribution data in accordance with a Bayesian deep neural network (B-DNN) architecture.
9. The AV system of claim 7 , wherein the one or more processors are configured to output the sets of distribution data in accordance with a non-Bayesian deep neural network (DNN) architecture.
10. The AV system of claim 7 , wherein the one or more processors are configured to output the sets of distribution data in accordance with a neural network architecture in which each one of the neural networks outputs a respective one of the sets of distribution data, and
wherein the one or more processors are configured to calculate respective uncertainty estimate values for each one of the sets of distribution data using predictive distributions obtained from multiple Monte Carlo forward passes through a respectively coupled one of the plurality of neural networks by sampling weights from a learned posterior distribution of each respectively coupled one of the plurality of neural networks.
11. The AV system of claim 7 , wherein the one or more processors are configured to monitor the uncertainty estimate values over time, and to flag an AV sensor as being faulty when a respective aleatoric uncertainty value associated with the AV sensor exceeds a threshold aleatoric uncertainty value.
12. The AV system of claim 7 , wherein the one or more processors are configured to output the sets of distribution data in accordance with a neural network architecture in which each one of the neural networks outputs a respective one of the sets of distribution data, and
wherein the one or more processors are configured to monitor the uncertainty estimate values over time, and to initiate a re-training sequence for the plurality of neural networks when an epistemic uncertainty value associated with one or more of the plurality of neural networks exceeds a threshold epistemic uncertainty value.
13. A non-transitory computer readable medium configured to store instructions that, when executed by one or more processors, cause the one or more processors to:
receive sensor data via each one of a respectively coupled autonomous vehicle (AV) sensor and output sets of distribution data, each one of the sets of distribution data representing a set of different classes used in accordance with an AV application environmental model;
calculate uncertainty estimate values comprising an aleatoric value and an epistemic uncertainty value for each one of the sets of distribution data;
gate a contribution of data provided by each AV sensor to each one of the sets of distribution data in accordance with a gating function that utilizes respectively calculated uncertainty estimate values for each one of a plurality of neural networks; and
generate the AV application environmental model using the gated distribution data.
14. The non-transitory computer readable medium of claim 13 , wherein the instructions include further instructions that, when executed by one or more processors, cause the one or more processors to output the sets of distribution data in accordance with a Bayesian deep neural network (B-DNN) architecture.
15. The non-transitory computer readable medium of claim 13 , wherein the instructions include further instructions that, when executed by one or more processors, cause the one or more processors to:
output the sets of distribution data in accordance with a neural network architecture in which each one of the neural networks outputs a respective one of the sets of distribution data, and
calculate respective uncertainty estimate values for each one of the sets of distribution data using predictive distributions obtained from multiple Monte Carlo forward passes through a respectively coupled one of the plurality of neural networks by sampling weights from a learned posterior distribution of each respectively coupled one of the plurality of neural networks.
16. The non-transitory computer readable medium of claim 13 , wherein the instructions include further instructions that, when executed by one or more processors, cause the one or more processors to monitor the uncertainty estimate values over time, and to flag an AV sensor as being faulty when a respective aleatoric uncertainty value associated with the AV sensor exceeds a threshold aleatoric uncertainty value.
17. The non-transitory computer readable medium of claim 13 , wherein the instructions include further instructions that, when executed by one or more processors, cause the one or more processors to:
output the sets of distribution data in accordance with a neural network architecture in which each one of the neural networks outputs a respective one of the sets of distribution data, and
monitor the uncertainty estimate values over time, and to initiate a re-training sequence for the plurality of neural networks when an epistemic uncertainty value associated with one or more of the plurality of neural networks exceeds a threshold epistemic uncertainty value.
18. The ECU of claim 1 , wherein the control unit is configured to generate the AV application environmental model using the gated distribution data such that inputs from AV sensors having lower aleatoric uncertainty values contribute more to predictions calculated by the AV application environmental model compared to AV sensors having higher aleatoric values.
19. The ECU of claim 1 , wherein the epistemic uncertainty value is with respect to a novelty of the distribution data output from a respectively coupled one of the plurality of neural networks.
20. The AV system of claim 7 , wherein the epistemic uncertainty value is with respect to a novelty of the distribution data output from a respectively coupled one of the plurality of neural networks.
21. The non-transitory computer readable medium of claim 13 , wherein the epistemic uncertainty value is with respect to a novelty of the distribution data output from a respectively coupled one of the plurality of neural networks.