Object detection and property determination for autonomous vehicles

1. A computer-implemented method of detecting objects, the method comprising:

obtaining, by a computing system comprising one or more computing devices, a representation of sensor data associated with an environment surrounding an autonomous vehicle, wherein the representation of sensor data comprises a plurality of sensor data points;

determining, by the computing system, a point classification and point property estimation for each of the plurality of the sensor data points;

clustering, by the computing system, a portion of the plurality of sensor data points into an object instance based at least in part on the point classification and point property estimation for each of the plurality of the sensor data points;

determining, by the computing system, a collection of point classifications and point property estimations for the portion of the plurality of sensor data points clustered into the object instance; and

determining, by the computing system, one or more object instance property estimations for the object instance based at least in part on the collection of point classifications and point property estimations for the portion of the plurality of sensor data points clustered into the object instance.

2. A computer-implemented method of claim 1 , wherein the one or more object instance property estimations are determined at least in part from a weighted average of the point property estimations for the portion of the plurality of sensor data points clustered into the object instance.

3. A computer-implemented method of claim 2 , further comprising:

determining, by the computing system, a weight for each sensor data point in the portion of the plurality of sensor data points clustered into the object instance that minimizes a variance of the weighted average of the point property estimations for the portion of the plurality of sensor data points clustered into the object instance.

4. A computer-implemented method of claim 1 , further comprising:

generating, by the computing system, a bounding shape estimation for the object instance based at least in part on the one or more object instance property estimations.

5. A computer-implemented method of claim 4 , further comprising:

determining, by the computing system, a bounding shape confidence score indicative of the predicted accuracy of the bounding shape.

6. A computer-implemented method of claim 1 , wherein obtaining, by the computing system, a representation of sensor data associated with an environment surrounding an autonomous vehicle comprises:

receiving, by the computing system, LIDAR data from one or more LIDAR sensors configured to transmit ranging signals relative to the autonomous vehicle; and

generating, by the computing system, a data matrix comprising a plurality of data channels based, at least in part, on the LIDAR data.

7. A computer-implemented method of claim 1 , further comprising:

accessing, by the computing system, a machine-learned detector model that has been trained to receive the representation of sensor data and to determine the object instance identified within the representation of sensor data; and wherein the one or more object instance property estimations are received as an output of the machine-learned detector model in response to receipt of the representation of sensor data as input.

8. A computer-implemented method of claim 7 , wherein the sensor data comprises LIDAR data, and wherein the machine-learned detector model comprises a convolutional neural network.

9. A computer-implemented method of claim 7 , wherein the machine-learned detector model has been trained in part by determining a loss function that computes error associated with a prediction made by the machine-learned detector model relative to the object instance.

10. A computer-implemented method of claim 1 , wherein the one or more object instance property estimations include one or more of an instance center, an orientation, and one or more dimensions of a bounding shape associated with the object instance.

11. An object detection system comprising:

one or more processors;

a machine-learned detector model configured to receive a representation of sensor data and, in response to receipt of the representation of sensor data, determine a point classification and point property estimation for a plurality of sensor data points, cluster a portion of the plurality of sensor data points into an object instance based at least in part on the point classification and point property estimation for the plurality of sensor data points, and output one or more object instance property estimations for the object instance that are determined at least in part from the point classifications and point property estimations for the portion of the plurality of sensor data points clustered into the object instance; and

at least one tangible, non-transitory computer readable medium that stores instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, the operations comprising:

obtaining a representation of sensor data associated with an environment surrounding an autonomous vehicle, wherein the representation of sensor data comprises a plurality of sensor data points;

inputting the representation of sensor data into the machine-learned detector model; and

receiving, as output of the machine-learned detector model, one or more object instance property estimations associated with an object instance detected within the representation of sensor data.

12. The object detection system of claim 11 , wherein the one or more object instance property estimations are determined at least in part from a weighted average of the point property estimations for the portion of the plurality of sensor data points clustered into the object instance.

13. The object detection system of claim 12 , wherein the machine-learned detector model is further configured to determine a weight for each sensor data point in the portion of the plurality of sensor data points clustered into the object instance that minimizes a variance of the weighted average of the point property estimations for the portion of the plurality of sensor data points clustered into the object instance.

14. The object detection system of claim 11 , wherein the representation of sensor data comprises a data matrix including a plurality of data channels that are determined at least in part from LIDAR data obtained from one or more LIDAR sensors configured to transmit ranging signals relative to the autonomous vehicle.

15. The object detection system of claim 11 , wherein the machine-learned detector model comprises a convolutional neural network.

16. The object detection system of claim 11 , wherein the machine-learned detector model has been trained in part by determining a loss function that computes error associated with a prediction made by the machine-learned detector model relative to the object instance as opposed to the respective sensor data points.

17. The object detection system of claim 11 , wherein the one or more object instance property estimations include one or more of an instance center, an orientation, and one or more dimensions of a bounding shape associated with the object instance.

18. An autonomous vehicle comprising:

a sensor system comprising at least one LIDAR sensor configured to transmit ranging signals relative to the autonomous vehicle and to generate LIDAR data; and

a vehicle computing system comprising:

one or more processors; and

at least one tangible, non-transitory computer readable medium that stores instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, the operations comprising:

receiving LIDAR data from the sensor system;

generating a data matrix comprising a plurality of data channels based at least in part on the LIDAR data;

providing the data matrix comprising a plurality of data channels as input to a machine-learned detector model configured to receive a data matrix of LIDAR data and, in response to receipt of the data matrix of LIDAR data, determine a point classification and point property estimation for a plurality of sensor data points, cluster a portion of the plurality of sensor data points into an object instance based at least in part on the point classification and point property estimation for the plurality of sensor data points, and output one or more object instance property estimations for the object instance that are determined at least in part from the point classifications and point property estimations for the portion of the plurality of sensor data points clustered into the object instance; and

receiving, as output of the machine-learned detector model, one or more object instance property estimations associated with an object instance detected within the representation of sensor data.

19. The autonomous vehicle of claim 18 , wherein the one or more object instance property estimations are determined at least in part from a weighted average of the point property estimations for the portion of the plurality of sensor data points clustered into the object instance.

20. The autonomous vehicle of claim 19 , wherein the machine-learned detector model is further configured to determine a weight for each sensor data point in the portion of the plurality of sensor data points clustered into the object instance that minimizes a variance of the weighted average of the point property estimations for the portion of the plurality of sensor data points clustered into the object instance.