Ensemble learning for cross-range 3D object detection in driver assist and autonomous driving systems
A cross-range 3D object detection method and system operable for training a 3D object detection model with N sub-groups of a point cloud corresponding to N detection distance ranges to form N 3D object detection models forming an ensemble 3D object detection model. Training the 3D object detection model with the N sub-groups of the point cloud corresponding to the N detection distance ranges includes training the 3D object detection model progressively from distant to near. Training the 3D object detection model with the N sub-groups of the point cloud corresponding to the N detection distance ranges includes, each time the 3D object detection model converges, saving resulting weights and adding a corresponding network to the ensemble 3D object detection model.
1. A cross-range 3D object detection method, comprising:
generating a point cloud only utilizing data indicated from a single measurement of a single optical device of a vehicle;
dividing the point cloud into N sub-groups corresponding to N detection distance ranges from the vehicle and defined with respect to the optical device, wherein N is an integer of three or more;
sequentially training a same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges from the vehicle and defined with respect to the optical device to form N 3D object detection algorithm versions corresponding to the N detection distance ranges utilizing stochastic gradient descent to converge and escape from a local minima of each of the N 3D object detection algorithm versions in combination with a cycling procedure in which a learning rate is abruptly raised and then quickly lowered to follow a cosine function and utilizing online ground truth boxes augmentation in which object boxes and inside points from one scene are copied to the same locations in another scene; and
combining the N 3D object detection algorithm versions trained form the same neural network to form an ensemble 3D object detection algorithm utilizing Snapshot Ensembling such that inference is applied to the N sub-groups of the point cloud by performing non-maximum suppression on predictions of all of the N 3D object detection algorithm versions.
2. The method of claim 1 , wherein sequentially training the same neural network for 3D object detection model with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises training the same neural network for 3D object detection progressively from distant to near.
3. The method of claim 1 , wherein sequentially training the same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises, each time the same neural network-converges, saving resulting weights and adding a corresponding network to the ensemble 3D object detection algorithm.
4. The method of claim 1 , wherein sequentially training the same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises training the same neural network separately for each of the N sub-groups.
5. The method of claim 1 , wherein sequentially training the same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises training the same neural network with the entire point cloud while progressively adding the N sub-groups from distant to near.
6. The method of claim 1 , wherein the 3D object detection algorithm versions each consist of a differently trained version of the same neural network for 3D object detection.
7. The method of claim 1 , wherein generating the point cloud only utilizing data indicated from the single measurement of the single optical device of the vehicle comprises generating the point cloud from a single image.
8. A cross-range 3D object detection system, comprising:
a memory storing instructions executed by a processor for:
generating a point cloud only utilizing data indicated from a single measurement of a single optical device of a vehicle;
dividing the point cloud into N sub-groups corresponding to N detection distance ranges from the vehicle and defined with respect to the optical device, wherein N is an integer of three or more;
sequentially training a same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges from the vehicle and defined with respect to the optical device to form N 3D object detection algorithm versions corresponding to the N detection distance ranges utilizing stochastic gradient descent to converge and escape from a local minima of each of the N 3D object detection algorithm versions in combination with a cycling procedure in which a learning rate is abruptly raised and then quickly lowered to follow a cosine function and utilizing online ground truth boxes augmentation in which object boxes and inside points from one scene are copied to the same locations in another scene; and
combining the N 3D object detection algorithm versions trained form the same neural network to form an ensemble 3D object detection algorithm utilizing Snapshot Ensembling such that inference is applied to the N sub-groups of the point cloud by performing non-maximum suppression on predications of all of the N 3D object detection algorithm versions.
9. The system of claim 8 , wherein sequentially training the same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises training the same neural network for 3D object detection progressively from distant to near.
10. The system of claim 8 , wherein sequentially training the same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises, each time the same neural network-converges, saving resulting weights and adding a corresponding network to the ensemble 3D object detection algorithm.
11. The system of claim 8 , wherein sequentially training the same neural network for 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises training the same neural network separately for each of the N sub-groups.
12. The system of claim 8 , wherein sequentially training the same neural network 3D object detection with the N sub-groups of the point cloud corresponding to the N detection distance ranges comprises training the same neural network with the entire point cloud while progressively adding the N sub-groups from distant to near.
13. The system of claim 8 , wherein the N 3D object detection algorithm versions each consist of a differently trained version of the same neural network for 3D object detection.