Distracted driving detection using a multi-task training process
Disclosed are a multi-task training technique and resulting model for detecting distracted driving. In one embodiment, a method is disclosed comprising inputting a plurality of labeled examples into a multi-task network, the multi-task network comprising: a backbone network, the backbone network generating one or more feature vectors corresponding to each of the labeled examples, and a plurality of prediction heads coupled to the backbone network; minimizing a joint loss based on outputs of the plurality of prediction heads, the minimizing the joint loss causing a change in parameters of the backbone network; and storing a distraction classification model after minimizing the joint loss, the distraction classification model comprising the parameters of the backbone network and parameters of at least one of the prediction heads.
1. A method comprising:
inputting a plurality of labeled examples into a multi-task network, the multi-task network comprising:
a backbone network comprising a convolutional neural network (CNN) and a feature pyramid network (FPN) coupled to the CNN, the backbone network generating one or more feature vectors corresponding to each of the labeled examples, and
a plurality of prediction heads coupled to the backbone network, wherein a subset of the plurality of prediction heads receives input from the CNN and a second subset of the plurality of prediction heads receives input from the FPN;
minimizing a joint loss based on outputs of the plurality of prediction heads, the minimizing the joint loss causing a change in parameters of the backbone network; and
storing a distraction classification model after minimizing the joint loss, the distraction classification model comprising the parameters of the backbone network and parameters of at least one of the prediction heads.
2. The method of claim 1 , wherein the CNN comprises an EfficientNet.
3. The method of claim 1 , wherein the FPN comprises a bi-directional FPN.
4. The method of claim 1 , wherein a subset of the plurality of prediction heads comprises a distraction classification prediction head, the distraction classification head comprising a convolutional layer, pooling layer, and fully connected layer.
5. The method of claim 1 , wherein a second subset of the plurality of prediction heads includes one or more of an object detection prediction head and a pose estimation prediction head.
6. The method of claim 5 , wherein the object detection prediction head comprises a bounding box regression network and an object class prediction network, each of the bounding box regression network and the object class prediction network comprising deep neural networks having a plurality of hidden layers, each hidden layer in the hidden layers comprising a convolutional layer, a batch normalization layer, and a batch activation layer, wherein the bounding box regression network outputs coordinates of a bounding box enclosing a detected object and the object class prediction network outputs a class corresponding to the detected object.
7. The method of claim 5 , wherein the pose estimation prediction head comprises a deep neural network, the deep neural network comprising a plurality of hidden layers and an output layer, each of the hidden layers comprising a convolutional layer, a batch normalization layer, and an activation layer, and the output layer comprising a convolutional layer.
8. The method of claim 1 , wherein storing a distraction classification model after minimizing the joint loss comprises storing parameters of the CNN and at least one of the prediction heads.
9. A non-transitory computer-readable storage medium for tangibly storing computer program instructions capable of being executed by a computer processor, the computer program instructions defining operations of:
inputting a plurality of labeled examples into a multi-task network, the multi-task network comprising:
a backbone network comprising a convolutional neural network (CNN) and a feature pyramid network (FPN) coupled to the CNN, the backbone network generating one or more feature vectors corresponding to each of the labeled examples, and
a plurality of prediction heads coupled to the backbone network, wherein a subset of the plurality of prediction heads receives input from the CNN and a second subset of the plurality of prediction heads receives input from the FPN;
minimizing a joint loss based on outputs of the plurality of prediction heads, the minimizing the joint loss causing a change in parameters of the backbone network; and
storing a distraction classification model after minimizing the joint loss, the distraction classification model comprising the parameters of the backbone network and parameters of at least one of the prediction heads.
10. The non-transitory computer-readable storage medium of claim 9 , wherein the CNN comprises an EfficientNet.
11. The non-transitory computer-readable storage medium of claim 9 , wherein the FPN comprises a bi-directional FPN.
12. The non-transitory computer-readable storage medium of claim 9 , wherein a subset of the plurality of prediction heads comprises a distraction classification prediction head, the distraction classification head comprising a convolutional layer, pooling layer, and fully connected layer.
13. The non-transitory computer-readable storage medium of claim 9 , wherein a second subset of the plurality of prediction heads includes one or more of an object detection prediction head and a pose estimation prediction head.
14. The non-transitory computer-readable storage medium of claim 9 , wherein storing a distraction classification model after minimizing the joint loss comprises storing parameters of the CNN and at least one of the prediction heads.
15. A device comprising:
a processor configured to:
input a plurality of labeled examples into a multi-task network, the multi-task network comprising: a backbone network comprising a convolutional neural network (CNN) and a feature pyramid network (FPN) coupled to the CNN, the backbone network generating one or more feature vectors corresponding to each of the labeled examples, and a plurality of prediction heads coupled to the backbone network, wherein a subset of the plurality of prediction heads receives input from the CNN and a second subset of the plurality of prediction heads receives input from the FPN,
minimize a joint loss based on outputs of the plurality of prediction heads, the minimizing the joint loss causing a change in parameters of the backbone network, and
store a distraction classification model after minimizing the joint loss, the distraction classification model comprising the parameters of the backbone network and parameters of at least one of the prediction heads.
16. The device of claim 15 , wherein storing a distraction classification model after minimizing the joint loss comprises storing parameters of the backbone network and at least one of the prediction heads.
17. The device of claim 15 , wherein the CNN comprises an EfficientNet.
18. The device of claim 15 , wherein the FPN comprises a bi-directional FPN.
19. The device of claim 15 , wherein a subset of the plurality of prediction heads comprises a distraction classification prediction head, the distraction classification head comprising a convolutional layer, pooling layer, and fully connected layer.
20. The device of claim 15 , wherein a second subset of the plurality of prediction heads includes one or more of an object detection prediction head and a pose estimation prediction head.