Generating training datasets for training machine learning based models for predicting behavior of traffic entities for navigating autonomous vehicles
A vehicle collects video data of an environment surrounding the vehicle including traffic entities, e.g., pedestrians, bicyclists, or other vehicles. The captured video data is sampled and the sampled video frames are presented to users to provide input on a traffic entity's state of mind. The system determines an attribute value that describes a statistical distribution of user responses for the traffic entity. If the attribute for a sampled video frame is within a threshold of the attribute of another video frame, the system interpolates attribute for a third video frame between the two sampled video frames. Otherwise, the system requests further user input for a video frame captured between the two sampled video frames. The interpolated and/or user based attributes are used to train a machine learning based model that predicts a hidden context of the traffic entity. The trained model is used for navigation of autonomous vehicles.
1. A method comprising:
receiving a sequence of video frames captured by a camera mounted on an autonomous vehicle;
sampling the sequence of video frames to obtain a subset of video frames;
annotating each of the subset of video frames obtained by sampling, each annotation specifying an attribute value describing a statistical distribution of user responses obtained by presenting the video frame to a plurality of users, each user response representing a likelihood of a state of mind represented by a hidden context for a traffic entity displayed in the video frame;
identifying a pair of video frames from the subset of video frames, the pair of video frames comprising a first video frame and a second video frame, wherein a time of capture of the first video frame and a time of capture of the second video frame is separated by a first time interval;
comparing a first attribute value associated with the first video frame and a second attribute value associated with the second video frame;
responsive to the first attribute value being within a threshold of the second attribute value, annotating a third video frame from the sequence of video frames having a time of capture within the first time interval by interpolating using the first attribute value and the second attribute value;
providing a training data set including the annotated subset of video frames and the third video frame for training a machine learning model, the machine learning model configured to receive an input video frame displaying a traffic entity and predict a statistical distribution of user responses representing the likelihood of the state of mind of the traffic entity displayed in the input video frame; and
providing the trained machine learning model to the autonomous vehicle to assist with navigation in traffic.
2. The method of claim 1 , further comprising:
identifying a second pair of video frames from the subset of video frames, the second pair of video frames comprising a fourth video frame and a fifth video frame, wherein the time of capture of the fourth video frame and the time of capture of the fifth video frame is separated by a second time interval;
comparing a fourth attribute value associated with the fourth video frame and a fifth attribute value associated with the fifth video frame; and
responsive to the fourth attribute value being greater than the threshold of the fifth attribute value:
identifying a sixth video frame from the sequence of video frames having a time of capture within the second time interval;
sending the sixth video frame to a plurality of users; and
annotating the sixth video frame based on responses from the plurality of users, the responses describing a hidden context associated with a traffic entity displayed in the sixth video frame.
3. The method of claim 2 , wherein the annotated sixth video frame is included in the training data set.
4. The method of claim 1 , wherein the state of mind of the traffic entity comprises an intention of the traffic entity to cross a path of the autonomous vehicle.
5. The method of claim 1 , wherein the state of mind of the traffic entity further comprises a measure of awareness of the autonomous vehicle.
6. The method of claim 1 , wherein an annotation of a video frame includes a set of parameters indicating a statistical distribution of the user responses provided by a set of users upon being presented with the video frame.
7. The method of claim 1 , wherein interpolating between the first attribute value and the second attribute value is based on the time of capture of the third video frame within the first time interval.
8. A computer readable non-transitory storage medium storing instructions, the instructions when executed by a processor cause the processor to perform steps comprising:
receiving a sequence of video frames captured by a camera mounted on an autonomous vehicle;
sampling the sequence of video frames to obtain subset of video frames;
annotating each of the subset of video frames obtained by sampling, each annotation specifying an attribute value describing a statistical distribution of user responses obtained by presenting the video frame to a plurality of users, each user response representing a likelihood of a state of mind represented by a hidden context for a traffic entity displayed in the video frame;
identifying a pair of video frames from the subset of video frames, the pair of video frames comprising a first video frame and a second video frame, wherein a time of capture of the first video frame and a time of capture of the second video frame is separated by a first time interval;
comparing a first attribute value associated with the first video frame and a second attribute value associated with the second video frame;
responsive to the first attribute value being within a threshold of the second attribute value, annotating a third video frame from the sequence of video frames having a time of capture within the first time interval by interpolating using the first attribute value and the second attribute value;
providing a training data set including the annotated subset of video frames and the third video frame for training a machine learning model, the machine learning model configured to receive an input video frame displaying a traffic entity and predict a statistical distribution of user responses representing the likelihood of the state of mind of the traffic entity displayed in the input video frame; and
providing the trained machine learning model to the autonomous vehicle to assist with navigation in traffic.
9. The computer readable non-transitory storage medium of claim 8 , wherein the instructions further cause the processor to perform steps comprising:
identifying a second pair of video frames from the subset of video frames, the second pair of video frames comprising a fourth video frame and a fifth video frame, wherein the time of capture of the fourth video frame and the time of capture of the fifth video frame is separated by a second time interval;
comparing a fourth attribute value associated with the fourth video frame and a fifth attribute value associated with the fifth video frame; and
responsive to the fourth attribute value being greater than the threshold of the fifth attribute value:
identifying a sixth video frame from the sequence of video frames having a time of capture within the second time interval;
sending the sixth video frame to a plurality of users; and
annotating the sixth video frame based on responses from the plurality of users, the responses describing a hidden context associated with a traffic entity displayed in the sixth video frame.
10. The computer readable non-transitory storage medium of claim 9 , wherein the annotated sixth video frame is included in the training data set.
11. The computer readable non-transitory storage medium of claim 9 , wherein the state of mind of the traffic entity comprises an intention of the traffic entity to cross a path of the autonomous vehicle.
12. The computer readable non-transitory storage medium of claim 9 , wherein the state of mind of the traffic entity further comprises a measure of awareness of the autonomous vehicle.
13. The computer readable non-transitory storage medium of claim 8 , wherein an annotation of a video frame includes a set of parameters indicating a statistical distribution of the user responses provided by a set of users upon being presented with the video frame.
14. The computer readable non-transitory storage medium of claim 8 , wherein interpolating between the first attribute value and the second attribute value is based on the time of capture of the third video frame within the first time interval.
15. A computer implemented system comprising:
a computer processor; and
a computer readable non-transitory storage medium storing instructions thereon, the instructions when executed by a processor cause the processor to perform steps of:
receiving a sequence of video frames captured by a camera mounted on an autonomous vehicle;
sampling the sequence of video frames to obtain subset of video frames;
annotating each of the subset of video frames obtained by sampling, each annotation specifying an attribute value describing a statistical distribution of user responses obtained by presenting the video frame to a plurality of users, each user response representing a likelihood of a state of mind represented by a hidden context for a traffic entity displayed in the video frame;
identifying a pair of video frames from the subset of video frames, the pair of video frames comprising a first video frame and a second video frame, wherein a time of capture of the first video frame and a time of capture of the second video frame is separated by a first time interval;
comparing a first attribute value associated with the first video frame and a second attribute value associated with the second video frame;
responsive to the first attribute value being within a threshold of the second attribute value, annotating a third video frame from the sequence of video frames having a time of capture within the first time interval by interpolating using the first attribute value and the second attribute value;
providing a training data set including the annotated subset of video frames and the third video frame for training a machine learning model, the machine learning model configured to receive an input video frame displaying a traffic entity and predict a statistical distribution of user responses representing the likelihood of the state of mind of the traffic entity displayed in the input video frame; and
providing the trained machine learning model to the autonomous vehicle to assist with navigation in traffic.
16. The computer system of claim 15 , wherein the instructions further cause the processor to perform steps comprising:
identifying a second pair of video frames from the subset of video frames, the second pair of video frames comprising a fourth video frame and a fifth video frame, wherein the time of capture of the fourth video frame and the time of capture of the fifth video frame is separated by a second time interval;
comparing a fourth attribute value associated with the fourth video frame and a fifth attribute value associated with the fifth video frame; and
responsive to the fourth attribute value being greater than the threshold of the fifth attribute value:
identifying a sixth video frame from the sequence of video frames having a time of capture within the second time interval;
sending the sixth video frame to a plurality of users; and
annotating the sixth video frame based on responses from the plurality of users, the responses describing a hidden context associated with a traffic entity displayed in the sixth video frame.
17. The computer system of claim 16 , wherein the annotated sixth video frame is included in the training data set.
18. The computer system of claim 15 , wherein interpolating between the first attribute value and the second attribute value is based on the time of capture of the third video frame within the first time interval.
19. The computer system of claim 15 , wherein the state of mind of the traffic entity comprises an intention of the traffic entity to cross a path of the autonomous vehicle.
20. The computer system of claim 15 , wherein the state of mind of the traffic entity further comprises a measure of awareness of the autonomous vehicle.