Multi-modal sensor data fusion for perception systems

1. A method of providing a perception system of a vehicle, the method comprising:

receiving multi-modal sensor data from a plurality of sensors mounted at different positions on the vehicle and having different modalities comprising an imaging sensor configured to capture a video input and a ranging sensor configured to capture a depth input;

deriving a plurality of horizontal and vertical components of optical flow data defining a change in scene dynamics between at least two frames of the video input based on a motion field change;

fusing multi-modal sensor data from the plurality of sensors having different modalities and the plurality of horizontal and vertical components of optical flow data;

detecting at least one region of interest in the multi-modal sensor data;

detecting one or more patches of interest in the multi-modal sensor data based on detecting the at least one region of interest;

applying a model comprising a deep convolutional neural network to the one or more patches of interest, wherein the model is trained using a supervised learning process for auto-registration of the multi-modal senor data and outputs classification values;

performing post-processing of a result of applying the model to produce a post-processing result for the one or more patches of interest, wherein the post-processing comprises aggregating the classification values and identifying a dominating label representing an alignment offset between at least two of the sensors to determine whether a misalignment exists for a same field of view; and

outputting a perception indication of the post-processing result to a decision-making process for autonomous navigation of a vehicle.

2. The method of claim 1 , wherein detecting the at least one region of interest comprises performing edge detection on one or more images in the multi-modal sensor data.

3. The method of claim 2 , wherein the edge detection comprises occlusion edge detection.

4. The method of claim 1 , wherein detecting the one or more patches of interest comprises applying a threshold on a percentage of pixels with edges in a given patch.

5. The method of claim 1 , wherein the perception indication is a misalignment indicator used to adjust at least one actuator of the at least two sensors having the misalignment.

6. The method of claim 1 , wherein the model is trained using a supervised learning process for occlusion edge labeling of the multi-modal senor data and outputs classification values.

7. The method of claim 6 , wherein the postprocessing comprises smoothing the classification, and further wherein the perception indication comprises identification of occlusion edges.

8. The method of claim 1 , further comprising:

adjusting at least two actuators in parallel based on the misalignment, each of the at least two actuators controlling one of the at least two sensors to correct the misalignment.

9. A method of providing a perception system of a vehicle, the method comprising:

receiving multi-modal sensor data from a plurality of sensors mounted at different positions on the vehicle and having different modalities comprising an imaging sensor configured to capture a video input and a ranging sensor configured to capture a depth input;

deriving a plurality of horizontal and vertical components of optical flow data defining a change in scene dynamics between at least two frames of the video input based on a motion field change;

fusing multi-modal sensor data from the plurality of sensors having different modalities and the plurality of horizontal and vertical components of optical flow data;

detecting at least one region of interest in the multi-modal sensor data;

detecting one or more patches of interest in the multi-modal sensor data based on detecting the at least one region of interest;

applying a model comprising a deep convolutional neural network to the one or more patches of interest, wherein the model is trained using an unsupervised learning process using a deep auto-encoder and outputs reconstruction errors;

performing post-processing of a result of applying the model to produce a post-processing result for the one or more patches of interest, wherein the post-processing comprises aggregating the reconstruction errors and identifying a misalignment between at least two of the sensors based on the aggregation of the reconstruction errors for a same field of view; and

outputting a perception indication of the post-processing result to a decision-making process for autonomous navigation of a vehicle.

10. The method of claim 9 , wherein the perception indication is a misalignment indicator used to adjust at least one actuator of the at least two sensors having the misalignment.

11. The method of claim 9 , further comprising:

adjusting at least two actuators in parallel based on the misalignment, each of the at least two actuators controlling one of the at least two sensors to correct the misalignment.

12. A system of a vehicle comprising:

a sensor system comprising a plurality of sensors mounted at different positions on the vehicle and having different modalities; and

a perception processor configured to:

receive multi-modal sensor data from a plurality of sensors having different modalities comprising an imaging sensor configured to capture a video input and a ranging sensor configured to capture a depth input;

derive a plurality of horizontal and vertical components of optical flow data defining a change in scene dynamics between at least two frames of the video input based on a motion field change;

fuse multi-modal sensor data from the plurality of sensors and the plurality of horizontal and vertical components of optical flow data;

detect at least one region of interest in the multi-modal sensor data;

detect one or more patches of interest in the multi-modal sensor data based on detecting the at least one region of interest, apply a model comprising a deep convolutional neural network to at least one detected patch of interest;

perform post-processing of a result of applying the model to produce a post-processing result for the at least one detected patch of interest; and

output a perception indication of the post-processing result to a decision-making process for autonomous navigation of a vehicle, wherein the model is trained using a supervised learning process for auto-registration of the multi-modal senor data and outputs classification values, and wherein the post-processing comprises aggregating the classification values and identifying a dominating label representing an alignment offset between at least two of the sensors to determine whether a misalignment exists of a same field of view.

13. The system of claim 12 , wherein the at least one region of interest is detected by performing edge detection on one or more images in the multi-modal sensor data.

14. The system of claim 12 , wherein detection of the one or more patches of interest comprises applying a threshold on a percentage of pixels with edges in a given patch.

15. The system of claim 12 , wherein the perception indication is a misalignment indicator used to adjust at least one actuator of the at least two sensors having the misalignment.

16. The system of claim 12 , wherein the model is trained using a supervised learning process for occlusion edge labeling of the multi-modal senor data and outputs classification values.

17. The system of claim 12 , wherein the postprocessing comprises aggregating the classification values and smoothing the classification, and further wherein the perception indication comprises identification of occlusion edges.

18. The system of claim 12 , further comprising:

at least two actuators, wherein the perception processor is configured to adjust the at least two actuators in parallel based on the misalignment, each of the at least two actuators controlling one of the at least two sensors to correct the misalignment.