Multiple resolution deep neural networks for vehicle autonomous driving systems

1. An autonomous driving system for a vehicle, the autonomous driving system comprising:

a sensor device configured to capture input data for an autonomous driving feature; and

a control system configured to:

obtain a training dataset for training a plurality of deep neural networks (DNNs) for the autonomous driving feature;

sub-sample the training dataset to obtain a plurality of training datasets comprising the training dataset and one or more sub-sampled training datasets each having a different resolution than a remainder of the plurality of training datasets;

train the plurality of DNNs using the plurality of training datasets, respectively;

determine, using the plurality of trained DNNs and the input data, a plurality of outputs for the autonomous driving feature; and

determine, using the plurality of outputs, a best output for the autonomous driving feature.

2. The autonomous driving system of claim 1 , wherein control system is configured to sub-sample the training dataset using sequential leader clustering.

3. The autonomous driving system of claim 2 , wherein the sequential lead clustering comprises, for each sub-sampled training dataset:

determining a different clustering threshold value;

for each sample of the training dataset, determining a distance between the sample and remainder of the samples of the training dataset; and

removing any of the plurality of samples of the training dataset that have a distance less than the clustering threshold value indicative of a match with another particular sample of the training dataset to generate the sub-sampled training dataset.

4. The autonomous driving system of claim 1 , wherein the control system is configured to train the plurality of DNNs using a different dropout rate for each of the plurality of DNNs.

5. The autonomous driving system of claim 4 , wherein the control system is further configured to:

determine a degradation state when the plurality of outputs continue to not differ by more than a threshold for a period of time; and

in response to determining the degradation state, at least one of set a fault and adjust or limit operation of the autonomous driving feature.

6. The autonomous driving system of claim 1 , wherein the control system is configured to determine the best output for the autonomous driving feature using a voting scheme.

7. The autonomous driving system of claim 1 , wherein the autonomous driving feature is object detection and tracking, wherein control system is configured to sub-sample the training dataset using sequential leader clustering, and wherein the sequential lead clustering comprises, for each sub-sampled training dataset:

(1) selecting a different clustering threshold (T);

(2) initializing a codebook (R) by randomly selecting a sample (x i ) from the training dataset (Ω);

(3) for each sample x i in Ω, compute a distance (d ij ) between x i and every sample x j in R;

(4) compute a minimum distance of d ij (d ij min), the distance between x i and a closest sample in R;

(5) when d ij min is smaller than T, the sample is not added to R;

(6) when d ij min is larger than T, the sample is added into the R; and

(7) output R as the sub-sampled training dataset.

8. The autonomous driving feature of claim 1 , wherein the sensor device is a camera system configured to capture images of an area external to the vehicle.

9. The autonomous driving system of claim 1 , wherein the sensor device is a light detection and ranging (LIDAR) system configured to capture LIDAR point cloud data for an area external to the vehicle.

10. A method for training multiple resolution deep neural networks (DNNs) for vehicle autonomous driving, the method comprising:

obtaining, by a control system, a training dataset for training a plurality of DNNs for an autonomous driving feature of the vehicle;

sub-sampling, by the control system, the training dataset to obtain a plurality of training datasets comprising the training dataset and one or more sub-sampled datasets each having a different resolution than a remainder of the plurality of training datasets;

training, by the control system, the plurality of DNNs using the plurality of training datasets, respectively;

determining, by the control system, a plurality of outputs for the autonomous driving feature using the plurality of trained DNNs and the input data;

receiving, by the control system, input data for the autonomous driving feature captured by a sensor device; and

determining, by the control system, a best output for the autonomous driving feature using the plurality of outputs.

11. The method of claim 10 , wherein the sub-sampling of the training dataset comprises using sequential leader clustering.

12. The method of claim 11 , wherein the sequential lead clustering comprises, for each sub-sampled training dataset:

determining, by the control system, a different clustering threshold value;

for each sample of the training dataset, determining, by the control system, a distance between the sample and remainder of the samples of the training dataset; and

removing, by the control system, any of the plurality of samples of the training dataset that have a distance less than the clustering threshold value indicative of a match with another particular sample of the training dataset to generate the sub-sampled training dataset.

13. The method of claim 10 , wherein training the plurality of DNNs comprises using a different dropout rate for each of the plurality of DNNs.

14. The method of claim 10 , wherein the determining of the best output for the autonomous driving feature comprises using a voting scheme.

15. The method of claim 14 , further comprising:

determining, by the control system, a degradation state when the plurality of outputs continue to not differ by more than a threshold for a period of time; and

in response to determining the degradation state, at least one of, by the control system, setting a fault and adjusting or limiting operation of the autonomous driving feature.

16. The method of claim 10 , wherein the autonomous driving feature is object detection and tracking, wherein the sub-sampling of the training dataset comprises using sequential leader clustering, and wherein the sequential lead clustering comprises, for each sub-sampled training dataset:

(1) selecting a different clustering threshold (T);

(2) initializing a codebook (R) by randomly selecting a sample (x i ) from the training dataset (Ω);

(3) for each sample x i in Ω, compute a distance (d ij ) between x i and every sample x j in R;

(4) compute a minimum distance of d ij (d ij min), the distance between x i and a closest sample in R;

(5) when d ij min is smaller than T, the sample is not added to R;

(6) when d ij min is larger than T, the sample is added into the R; and

(7) output R as the sub-sampled training dataset.

17. The method of claim 10 , wherein the sensor device is a camera system configured to capture images of an area external to the vehicle.

18. The method of claim 10 , wherein the sensor device is a light detection and ranging (LIDAR) system configured to capture LIDAR point cloud data for an area external to the vehicle.