Occlusion-aware prediction of human behavior
An occlusion analysis system improves accuracy of behavior prediction models by generating occlusion parameters that may inform mathematical models to generate more accurate predictions. The occlusion analysis system trains and applies models for generating occlusion parameters, such as a manner in which a person is occluded, occlusion percentage, occlusion type. A behavior prediction system may input the occlusion parameters as well as other parameters relating to activity of the human into a second mathematical model for behavior prediction. The second machine learning model is a higher-level model trained to output a prediction that the human will exhibit a future behavior and a confidence level associated with the prediction. The confidence level is at least partially determined based on the occlusion parameters. The behavior prediction system may output the prediction and the confidence level to a control system that generates commands associated with a vehicle and other intelligent video analytics systems.
1. A method comprising:
receiving a video comprising a sequence of frames where at least a part of a human is visible in a given frame of the sequence of frames;
applying the sequence of frames to an occlusion model, wherein the occlusion model is a machine learning model trained to detect occlusion of the human in the video;
receiving, as output from the occlusion model, one or more occlusion parameters;
inputting the one or more occlusion parameters as well as other parameters relating to activity of the human into a second model that predicts that the human will exhibit a future behavior or recognizes a current behavior of the human based on the other parameters, wherein the second model outputs a confidence level associated with the prediction that is based on the one or more occlusion parameters, wherein the confidence level is determined based on whether the one or more occlusion parameters indicate that the human is occluded in the given frame; and
outputting the prediction and the confidence level to a control system, wherein the control system, responsive to the confidence level being below a first threshold level associated with the second model, selects a third model for behavior predictions, wherein a predicted confidence level for the third model predicted using the one or more occlusion parameters is greater than a second threshold confidence level associated with the third model.
2. The method of claim 1 , wherein the one or more occlusion parameters in the output from the occlusion model is a percentage of the human that is occluded in the given frame.
3. The method of claim 2 , wherein the one or more occlusion parameters in the output from the occlusion model include a degree of direction of an occlusion of the human in the given frame.
4. The method of claim 3 , wherein the percentage and the degree of direction of occlusion are used to adjust predicted velocity and distance of the human.
5. The method of claim 1 , wherein the one or more occlusion parameters includes a binary indication of whether the human is occluded, and wherein the binary indication is used as input for the second model.
6. The method of claim 1 , wherein the occlusion model predicts a second bounding box illustrating the human without occlusion based on a first bounding box depicting the part of the human that is not occluded.
7. The method of claim 1 , wherein the occlusion model predicts a pose of the human based on a part of the human in the given frame that is not occluded.
8. The method of claim 1 further comprising, responsive to detecting occlusion in the given frame, determining to use another given frame that is before or after the given frame in the video, wherein the human is not occluded in the other given frame.
9. The method of claim 1 , further comprising:
responsive to detecting that processing power is below a threshold, determining to focus the processing power on an area of focus of the given frame based on the one or more occlusion parameters.
10. The method of claim 1 , wherein the occlusion model further detects occlusion caused by a lighting environment.
11. The method of claim 1 , wherein the occlusion model determines that the human belongs to a group of people based on an analysis of a pattern that the human is occluded by another human in the group.
12. The method of claim 1 , wherein the occlusion model is a standalone model or is embedded in the second model.
13. A non-transitory computer-readable storage medium comprising memory with executable computer instructions encoded thereon that, when executed by one or more processors, cause the one or more processors to perform operations, the instructions comprising instructions to:
receive a video comprising a sequence of frames where at least a part of a human is visible in a given frame of the sequence of frames;
apply the sequence of frames to an occlusion model, wherein the occlusion model is a machine learning model trained to detect occlusion of the human in the video;
receive, as output from the occlusion model, one or more occlusion parameters;
input the one or more occlusion parameters as well as other parameters relating to activity of the human into a second model that predicts that the human will exhibit a future behavior or recognizes a current behavior of the human based on the other parameters, and to also output a confidence level associated with the prediction that is based on the one or more occlusion parameters, wherein the confidence level is determined based on whether the one or more occlusion parameters indicate that the human is occluded in the given frame; and
output the prediction and the confidence level to a control system, wherein the control system, responsive to the confidence level being below a first threshold level associated with the second model, selects a third model for behavior predictions, wherein a predicted confidence level for the third model predicted using the one or more occlusion parameters is greater than a second threshold confidence level associated with the third model.
14. The non-transitory computer-readable storage medium of claim 13 , wherein the one or more occlusion parameters in the output from the occlusion model is a percentage of the human that is occluded in the given frame.
15. The non-transitory computer-readable storage medium of claim 13 , wherein the one or more occlusion parameters includes a binary indication of whether the human is occluded, and wherein the binary indication is used as input for the second model.
16. The non-transitory computer-readable storage medium of claim 13 , wherein the instructions further comprise instructions to, responsive to detecting occlusion in the given frame, determine to use another given frame that is before or after the given frame in the video, wherein the human is not occluded in the other given frame.