Computer-implemented method and system for generating synthetic sensor data, and training method
A method generates synthetic sensor data corresponding to a LiDAR sensor of a vehicle, the synthetic sensor data including superimposed distance and intensity information. The method includes: providing a hierarchical variational autoencoder; conditioning a first feature vector and a second feature vector with a second data set, the second data set including distance and intensity information; combining the conditioned first feature vector and the conditioned second feature vector into a resulting third feature vector; and decoding the resulting third feature vector to generate a third data set of synthetic sensor data, the third data set including superimposed distance and intensity information.
1. A method for generating synthetic sensor data corresponding to a LiDAR sensor of a vehicle, the synthetic sensor data including superimposed distance and intensity information, the method comprising:
providing a hierarchical variational autoencoder, wherein the hierarchical variational autoencoder has a first level of hierarchy and a second level of hierarchy, and wherein the hierarchical variational autoencoder has a third level of hierarchy or is configured to communicate with a third level of hierarchy of an external variational autoencoder;
receiving, by a variational autoencoder of the first level of hierarchy, a first data set of LiDAR sensor data including distance information, wherein the first data set comprises synthetically generated and/or captured real sensor data, the variational autoencoder of the first level of hierarchy assigning global features of the first data set to a first codebook vector;
receiving, by a variational autoencoder of the second level of hierarchy, the first data set, the variational autoencoder of the second level of hierarchy assigning local features of the first data set to a second codebook vector;
conditioning a first feature vector encoded by the variational autoencoder of the first level of hierarchy and a second feature vector encoded by the variational autoencoder of the second level of hierarchy with a second data set of LiDAR sensor data from the LiDAR sensor of the vehicle, the second data set including distance and intensity information;
combining the conditioned first feature vector and the conditioned second feature vector into a resulting third feature vector; and
decoding the resulting third feature vector to generate a third data set of synthetic LiDAR sensor data, the third data set including superimposed distance and intensity information;
wherein the first data set is encoded by a first encoder of the hierarchical variational autoencoder, wherein the encoding by the first encoder reduces an image resolution of the first data set;
wherein the first data set encoded by the first encoder is divided into the first level of hierarchy and the second level of hierarchy, wherein the first level of hierarchy of the first data set is further encoded by a second encoder of the hierarchical variational autoencoder, and wherein the encoding by the second encoder reduces the image resolution of the first data set;
wherein the first data set is encoded into the first feature vector by the second encoder. and wherein the first feature vector is assigned to the first codebook vector, which has the smallest distance to the first feature vector, by a first artificial convolutional neural network of the first level of hierarchy.
2. The method according to claim 1 , wherein the first codebook vector is decoded by a first decoder of the hierarchical variational autoencoder, wherein the decoding by the first decoder increases an image resolution of the first codebook vector.
3. The method according to claim 2 , wherein the first data set is output by the first decoder, and wherein the first data set output by the first decoder and the first data set encoded by the first encoder of the hierarchical variational autoencoder are combined into the resulting third feature vector in the second level of hierarchy.
4. The method according to claim 3 , wherein the resulting third feature vector is assigned to the second codebook vector, which has the smallest distance to the resulting third feature vector, by a second artificial convolutional neural network of the second level of hierarchy.
5. The method according to claim 1 , wherein the second data set from the LiDAR sensor of the vehicle is encoded by a third encoder in the third level of hierarchy of the hierarchical variational autoencoder or the external variational autoencoder, wherein the encoding by the third encoder reduces an image resolution of the second data set.
6. The method according to claim 5 , wherein the second data set is encoded into a fourth feature vector by the third encoder, and wherein the fourth feature vector is assigned to a third codebook vector, which has the smallest distance to the fourth feature vector of the second data set, by a third artificial convolutional neural network of the third level of hierarchy of the hierarchical variational autoencoder or the external variational autoencoder.
7. The method according to claim 6 , wherein the third codebook vector is decoded by a second decoder of the hierarchical variational autoencoder or the external variational autoencoder, wherein the second data set is output by the second decoder, and wherein the second data set conditions the first feature vector encoded by the variational autoencoder of the first level of hierarchy and the second feature vector encoded by the variational autoencoder of the second level of hierarchy.
8. The method according to claim 1 , wherein the first feature vector encoded by the variational autoencoder of the first level of hierarchy and the second feature vector encoded by the variational autoencoder of the second level of hierarchy are conditioned with a data set label, the data set label indicating whether corresponding sensor data is synthetically generated or captured real sensor data.
9. A non-transitory computer-readable medium having processor-executable instructions stored thereon for generating synthetic sensor data corresponding to a LiDAR sensor of a vehicle, the synthetic sensor data including superimposed distance and intensity information, wherein the processor-executable instructions, when executed, facilitate performance of the following:
providing a hierarchical variational autoencoder, wherein the hierarchical variational autoencoder has a first level of hierarchy and a second level of hierarchy, and wherein the hierarchical variational autoencoder has a third level of hierarchy or is configured to communicate with a third level of hierarchy of an external variational autoencoder;
receiving, by a variational autoencoder of the first level of hierarchy, a first data set of LiDAR sensor data including distance information, wherein the first data set comprises synthetically generated and/or captured real sensor data, the variational autoencoder of the first level of hierarchy assigning global features of the first data set to a first codebook vector;
receiving, by a variational autoencoder of the second level of hierarchy, the first data set, the variational autoencoder of the second level of hierarchy assigning local features of the first data set to a second codebook vector;
conditioning a first feature vector encoded by the variational autoencoder of the first level of hierarchy and a second feature vector encoded by the variational autoencoder of the second level of hierarchy with a second data set of LiDAR sensor data from the LiDAR sensor of the vehicle, the second data set including distance and intensity information;
combining the conditioned first feature vector and the conditioned second feature vector into a resulting third feature vector; and
decoding the resulting third feature vector to generate a third data set of synthetic LiDAR sensor data, the third data set including superimposed distance and intensity information;
wherein the first data set is encoded by a first encoder of the hierarchical variational autoencoder, wherein the encoding by the first encoder reduces an image resolution of the first data set;
wherein the first data set encoded by the first encoder is divided into the first level of hierarchy and the second level of hierarchy, wherein the first level of hierarchy of the first data set is further encoded by a second encoder of the hierarchical variational autoencoder. and wherein the encoding by the second encoder reduces the image resolution of the first data set;
wherein the first data set is encoded into the first feature vector by the second encoder. and wherein the first feature vector is assigned to the first codebook vector, which has the smallest distance to the first feature vector, by a first artificial convolutional neural network of the first level of hierarchy.
10. A method for generating synthetic sensor data corresponding to a LiDAR sensor of a vehicle, the synthetic sensor data including superimposed distance and intensity information, the method comprising:
providing a hierarchical variational autoencoder, wherein the hierarchical variational autoencoder has a first level of hierarchy and a second level of hierarchy, and wherein the hierarchical variational autoencoder has a third level of hierarchy or is configured to communicate with a third level of hierarchy of an external variational autoencoder;
receiving, by a variational autoencoder of the first level of hierarchy, a first data set of LiDAR sensor data including distance information, wherein the first data set comprises synthetically generated and/or captured real sensor data, the variational autoencoder of the first level of hierarchy assigning global features of the first data set to a first codebook vector;
receiving, by a variational autoencoder of the second level of hierarchy, the first data set, the variational autoencoder of the second level of hierarchy assigning local features of the first data set to a second codebook vector;
conditioning a first feature vector encoded by the variational autoencoder of the first level of hierarchy and a second feature vector encoded by the variational autoencoder of the second level of hierarchy with a second data set of LiDAR sensor data from the LiDAR sensor of the vehicle, the second data set including distance and intensity information;
combining the conditioned first feature vector and the conditioned second feature vector into a resulting third feature vector; and
decoding the resulting third feature vector to generate a third data set of synthetic LiDAR sensor data, the third data set including superimposed distance and intensity information;
wherein the second data set from the LiDAR sensor of the vehicle is encoded by a third encoder in the third level of hierarchy of the hierarchical variational autoencoder or the external variational autoencoder, wherein the encoding by the third encoder reduces an image resolution of the second data set;
wherein the second data set is encoded into a fourth feature vector by the third encoder, and wherein the fourth feature vector is assigned to a third codebook vector, which has the smallest distance to the fourth feature vector of the second data set, by a third artificial convolutional neural network of the third level of hierarchy of the hierarchical variational autoencoder or the external variational autoencoder.