Multi-axis radar velocity image from stereo doppler radar

1 . A method comprising:

receiving, at a computing system, first radar data from a first radar and second radar data from a second radar, wherein the first radar and the second radar are coupled at different locations on a vehicle traveling in an environment;

determining a first radial speed for an object in the environment based on the first radar data and a second radial speed for the object based on the second radar data;

estimating, based on the first radial speed and the second radial speed, a velocity vector for the object relative to the vehicle;

generating radar imagery based on the first radar data and the second radar data, wherein a first channel of the radar imagery represents x-component velocity information of the estimated velocity vector on a per pixel basis and a second channel of the radar imagery represents y-component velocity information of the estimated velocity vector on the per pixel basis;

providing the radar imagery as an input into a neural network; and

controlling the vehicle based on an output from the neural network.

2 . The method of claim 1 , further comprising:

transforming the first radial speed from a first frame corresponding to the first radar into a first vector in a vehicle frame corresponding to the vehicle; and

transforming the second radial speed from a second frame corresponding to the second radar into a second vector in the vehicle frame.

3 . The method of claim 2 , wherein estimating the velocity vector for the object comprises:

determining an intersection point between a first line extending orthogonal to the first vector and a second line extending orthogonal to the second vector; and

estimating the velocity vector for the object based on the determined intersection point and a position of the object.

4 . The method of claim 3 , further comprising:

receiving third radar data from a third radar, wherein the third radar is coupled to the vehicle at a third location that differs from locations of the first radar and the second radar;

determining a third radial speed for the object in the environment based on the third radar data; and

wherein estimating the velocity vector for the object relative to the vehicle comprises:

estimating the velocity vector further based on the third radial speed for the object.

5 . The method of claim 4 , further comprising:

transforming the third radial speed into a third vector in the vehicle frame; and

determining an intersection point between a first plane extending orthogonal to the first vector, a second plane extending orthogonal to the second vector, and a third plane extending orthogonal to the third vector; and

estimating the velocity vector for the object based on the determined intersection point and a position of the object.

6 . The method of claim 5 , wherein determining the intersection point between the first plane, the second plane, and the third plane comprises:

solving a three by three linear system of equations generated based on the intersection point between the first plane, the second plane, and the third plane.

7 . The method of claim 1 , wherein controlling the vehicle based on the output from the neural network comprises:

controlling the vehicle based on object attributes output by the neural network for the object.

8 . The method of claim 7 , wherein the object attributes indicate one or more of a pose, a position, a type, and an intent for the object.

9 . The method of claim 1 , further comprising:

providing images from a camera in addition to the radar imagery as inputs into the neural network.

10 . The method of claim 1 , wherein the neural network is configured to initially track the object based on the radar imagery.

11 . The method of claim 1 , further comprising:

causing the first radar to transmit first radar signals into the environment based on a first waveform and a first pulse repetition frequency; and

causing the second radar to transmit second radar signals into the environment based on a second waveform and a second pulse repetition frequency.

12 . The method of claim 1 , further comprising:

providing, based on the first radar data and the second radar data, a return intensity as a second input into the neural network.

13 . The method of claim 12 , wherein providing the return intensity as the second input into the neural network comprises:

determining a mean return intensity based on a first return intensity measured for the object in the first radar data and a second return intensity measured for the object in the second radar data; and

providing the mean return intensity as the second input into the neural network.

14 . The method of claim 12 , wherein providing the return intensity as the second input into the neural network comprises:

comparing a first max return intensity measured for the object in the first radar data and a second max return intensity measured for the object in the second radar data;

determining that the first max return intensity is greater than the second max return intensity; and

based on determining that the first max return intensity is greater than the second max return intensity, providing the first max return intensity as the second input into the neural network.

15 . The method of claim 1 , further comprising:

aligning pixels of a first radar image generated based on the first radar data and a second radar image generated based on the second radar data into a common coordinate frame using an extrinsic transform between the first radar and the second radar.

16 . The method of claim 1 , wherein the first radar and the second radar have partially overlapping fields of view, and wherein the radar imagery is 2-channel radar imagery.

17 . The method of claim 1 , wherein the neural network performs object detection based on the radar imagery without a Kalman filter.

18 . A system comprising:

a vehicle;

a first radar and a second radar, wherein the first radar and the second radar are coupled at different locations on the vehicle;

a computing device configured to:

receive first radar data from a first radar and second radar data from a second radar;

determine a first radial speed for an object in an environment of the vehicle based on the first radar data and a second radial speed for the object based on the second radar data;

estimate, based on the first radial speed and the second radial speed, a velocity vector for the object relative to the vehicle;

generate radar imagery based on the first radar data and the second radar data, wherein a first channel of the radar imagery represents x-component velocity information of the estimated velocity vector on a per pixel basis and a second channel of the radar imagery represents y-component velocity information of the estimated velocity vector on the per pixel basis;

provide the radar imagery as an input into a neural network; and

control the vehicle based on an output from the neural network.

19 . The system of claim 18 , wherein the neural network is configured to use lidar point cloud data and images as additional inputs.

20 . A non-transitory computer-readable medium configured to store instructions, that when executed by a computing system, causes the computing system to perform operations comprising:

receiving first radar data from a first radar and second radar data from a second radar, wherein the first radar and the second radar are coupled at different locations on a vehicle traveling in an environment;

determining a first radial speed for an object in the environment based on the first radar data and a second radial speed for the object based on the second radar data;

estimating, based on the first radial speed and the second radial speed, a velocity vector for the object relative to the vehicle;

generating radar imagery based on the first radar data and the second radar data, wherein a first channel of the radar imagery represents x-component velocity information of the estimated velocity vector on a per pixel basis and a second channel of the radar imagery represents y-component velocity information of the estimated velocity vector on the per pixel basis;

providing the radar imagery as an input into a neural network; and

controlling the vehicle based on an output from the neural network.