Modular predictions for complex human behaviors
A device performs operations including determining a probability that a vulnerable road user (VRU) will continue on a current path (e.g., in connection with controlling an autonomous vehicle). The device receives an image depicting a vulnerable road user (VRU). The device inputs at least a portion of the image into a model, and receives, as output from the model, a plurality of probabilities describing the VRU, each of the probabilities corresponding to a probability that the VRU is in a given state. The device determines, based on at least some of the plurality of probabilities, a probability that the VRU will exhibit a behavior, and outputs the probability that the VRU will exhibit the behavior to a control system.
1. A method comprising:
receiving a plurality of sequential images comprising a first image depicting a human captured at a first time and a second image depicting the human captured at a second time later than the first time;
inputting at least a portion of the first image into a hybrid model comprising a deep learning model and a probabilistic graphical model, the deep learning model comprising a multi-task model having different branches, each different branch trained to determine a different feature;
receiving, as output from the hybrid model, a plurality of probabilities corresponding to a probability for a given variable feature corresponding to different states of the human;
inputting the plurality of probabilities and the second image into the hybrid model and receiving, as output from the hybrid model, a confidence value that the human will exhibit a behavior; and
outputting the confidence value that the human will exhibit the behavior to a control system.
2. The method of claim 1 , wherein the behavior is continuing on a current path, and wherein the method further comprises:
determining, based on at least two or more of the plurality of probabilities, a probability that the human is distracted, wherein determining the probability that the human will continue on the current path is based on the probability that the human is distracted.
3. The method of claim 2 , wherein determining, based on the at least two or more of the plurality of probabilities, the probability that the human is distracted comprises:
determining that a given probability of the plurality of probabilities has a conditional independency, where the given probability exceeds a threshold; and
responsive to determining that the given probability exceeds the threshold, determining the probability that the human is distracted by using the given probability to an exclusion of at least one other probability of the plurality of probabilities.
4. The method of claim 3 , further comprising:
determining the at least one other probability of the plurality of probabilities that is excluded based on a data structure that defines the conditional independency.
5. The method of claim 1 , wherein the probabilistic graphical model is at least one of a Markov network and a Bayesian network.
6. The method of claim 1 , wherein the confidence value that the human will exhibit the behavior is determined based on a posterior predictive distribution of possible outcomes, and wherein the confidence value is determined based on a spread of the posterior predictive distribution.
7. The method of claim 1 , wherein the control system determines whether to issue a control signal based on the confidence value score.
8. The method of claim 1 , wherein the hybrid model is applicable in a plurality of domains without a need for domain-specific training data.
9. A non-transitory computer-readable medium comprising memory with instructions encoded thereon, the instructions causing one or more processors to perform operations when executed, the instructions comprising instructions to:
receiving a plurality of sequential images comprising a first image depicting a human captured at a first time and a second image depicting the human captured at a second time later than the first time;
inputting at least a portion of the first image into a hybrid model comprising a deep learning model and a probabilistic graphical model, the deep learning model comprising a multi-task model having different branches, each different branch trained to determine a different feature;
receiving, as output from the hybrid model, a plurality of probabilities corresponding to a probability for a given variable feature corresponding to different states of the human;
inputting the plurality of probabilities and the second image into the hybrid model and receiving, as output from the hybrid model, a confidence value that the human will exhibit a behavior; and
outputting the confidence value that the human will exhibit the behavior to a control system.
10. The non-transitory computer-readable medium of claim 9 , wherein the behavior is continuing on a current path, and wherein the instructions further comprise instructions to:
determine, based on at least two or more of the plurality of probabilities, a probability that the human is distracted, wherein determining the probability that the human will continue on the current path is based on the probability that the human is distracted.
11. The non-transitory computer-readable medium of claim 10 , wherein the instructions to determine, based on the at least two or more of the plurality of probabilities, the probability that the human is distracted comprise instructions to:
determine that a given probability of the plurality of probabilities has a conditional independency, where the given probability exceeds a threshold; and
responsive to determining that the given probability exceeds the threshold, determine the probability that the human is distracted by using the given probability to an exclusion of at least one other probability of the plurality of probabilities.
12. The non-transitory computer-readable medium of claim 11 , the instructions further comprising instructions to:
determine the at least one other probability of the plurality of probabilities that is excluded based on a data structure that defines the conditional independency.
13. The non-transitory computer-readable medium of claim 9 , wherein the probabilistic graphical model is at least one of a Markov network and a Bayesian network.
14. The non-transitory computer-readable medium of claim 9 , wherein the confidence value that the human will exhibit the behavior is determined based on a posterior predictive distribution of possible outcomes, and wherein the confidence value is determined based on a spread of the posterior predictive distribution.
15. The non-transitory computer-readable medium of claim 9 , wherein the control system determines whether to issue a control signal based on the confidence value.
16. The non-transitory computer-readable medium of claim 9 , wherein the hybrid model is applicable in a plurality of domains without a need for domain-specific training data.
17. A system comprising:
memory with instructions encoded thereon; and
one or more processors that, when executing the instructions, are caused to perform operations comprising:
receiving a plurality of sequential images comprising a first image depicting a human captured at a first time and a second image depicting the human captured at a second time later than the first time;
inputting at least a portion of the first image into a hybrid model comprising a deep learning model and a probabilistic graphical model, the deep learning model comprising a multi-task model having different branches, each different branch trained to determine a different feature;
receiving, as output from the hybrid model, a plurality of probabilities corresponding to a probability for a given variable feature corresponding to different states of the human;
inputting the plurality of probabilities and the second image into the hybrid model and receiving, as output from the hybrid model, a confidence value that the human will exhibit a behavior; and
outputting the confidence value that the human will exhibit the behavior to a control system.
18. The system of claim 17 , wherein the behavior is continuing on a current path, and wherein the operations further comprise:
determining, based on at least two or more of the plurality of probabilities, a probability that the human is distracted, wherein determining the probability that the human will continue on the current path is based on the probability that the human is distracted.
19. The system of claim 18 , wherein determining, based on the at least two or more of the plurality of probabilities, the probability that the human is distracted comprises:
determining that a given probability of the plurality of probabilities has a conditional independency, where the given probability exceeds a threshold; and
responsive to determining that the given probability exceeds the threshold, determining the probability that the human is distracted by using the given probability to an exclusion of at least one other probability of the plurality of probabilities.
20. The system of claim 19 , further comprising:
determining the at least one other probability of the plurality of probabilities that is excluded based on a data structure that defines the conditional independency.