Lane change prediction on highways
Disclosed are techniques for drive trajectory prediction. In one or more aspects, an ego vehicle applies a machine learning model to one or more agent tensors and one or more map tensors associated with a target vehicle to obtain a lane change classification label representing a predicted lane change intention of the target vehicle on a multi-lane highway, wherein the predicted lane change intention includes at least a lane change classification, and performs a driving maneuver based on the predicted lane change intention.
1 . An ego vehicle, comprising:
one or more memories;
one or more transceivers; and
one or more processors communicatively coupled to the one or more memories and the one or more transceivers, the one or more processors, either alone or in combination, configured to:
apply an encoder-decoder machine learning model to one or more agent tensors and one or more map tensors associated with a target vehicle to obtain a lane change classification label representing a predicted lane change intention of the target vehicle on a multi-lane highway, wherein the predicted lane change intention includes at least a lane change classification, and wherein a decoder side of the encoder-decoder machine learning model outputs the lane change classification label; and
perform a driving maneuver based on the predicted lane change intention.
2 . The ego vehicle of claim 1 , wherein the one or more agent tensors represent a plurality of features of at least the target vehicle and one or more neighbor vehicles of the target vehicle over a most recent period of time.
3 . The ego vehicle of claim 2 , wherein the plurality of features comprises:
current x coordinates of the target vehicle and the one or more neighbor vehicles,
current y coordinates of the target vehicle and the one or more neighbor vehicles,
previous x coordinates of the target vehicle and the one or more neighbor vehicles,
previous y coordinates of the target vehicle and the one or more neighbor vehicles,
x-axis velocity values of the target vehicle and the one or more neighbor vehicles,
y-axis velocity values of the target vehicle and the one or more neighbor vehicles,
angular velocity values of the target vehicle and the one or more neighbor vehicles,
acceleration values of the target vehicle and the one or more neighbor vehicles,
blinker states of the target vehicle and the one or more neighbor vehicles,
a time offset between current and previous positions of the target vehicle and the one or more neighbor vehicles,
flags indicating a presence of the target vehicle and the one or more neighbor vehicles, or
any combination thereof.
4 . The ego vehicle of claim 3 , wherein a length of the most recent period of time is one second.
5 . The ego vehicle of claim 1 , wherein the one or more map tensors represent map context information around the target vehicle.
6 . The ego vehicle of claim 5 , wherein the map context information comprises:
map lane center points for a current lane of the target vehicle,
map lane center points for a lane to the left of the target vehicle,
map lane center points for a lane to the right of the target vehicle, or
any combination thereof.
7 . The ego vehicle of claim 1 , wherein an encoder side of the encoder-decoder machine learning model comprises:
an agent encoding module,
a map encoding module, and
an aggregation module.
8 . The ego vehicle of claim 7 , wherein the agent encoding module comprises a self-attention module.
9 . The ego vehicle of claim 7 , wherein the aggregation module aggregates results of the agent encoding module and the map encoding module.
10 . The ego vehicle of claim 7 , wherein the aggregation module comprises a map attention module and a cross attention module.
11 . The ego vehicle of claim 10 , wherein:
the aggregation module further comprises a multi-layer perception (MLP) layer, and
the MLP layer outputs results of the aggregation module.
12 . The ego vehicle of claim 1 , wherein the decoder side of the encoder-decoder machine learning model comprises:
one or more MLP layers that output the lane change classification label.
13 . The ego vehicle of claim 1 , wherein the predicted lane change intention further includes one or more predicted lane change trajectories associated with the lane change classification.
14 . The ego vehicle of claim 1 , wherein the lane change classification label represents one of:
a lane change left,
a lane change right, or
no lane change.
15 . The ego vehicle of claim 1 , wherein the driving maneuver comprises:
a lane change left,
a lane change right, or
no lane change.
16 . A method of drive trajectory prediction performed by an ego vehicle, comprising:
applying an encoder-decoder machine learning model to one or more agent tensors and one or more map tensors associated with a target vehicle to obtain a lane change classification label representing a predicted lane change intention of the target vehicle on a multi-lane highway, wherein the predicted lane change intention includes at least a lane change classification, and wherein a decoder side of the encoder-decoder machine learning model outputs the lane change classification label; and
performing a driving maneuver based on the predicted lane change intention.
17 . The method of claim 16 , wherein the one or more agent tensors represent a plurality of features of at least the target vehicle and one or more neighbor vehicles of the target vehicle over a most recent period of time.
18 . The method of claim 17 , wherein the plurality of features comprises:
current x coordinates of the target vehicle and the one or more neighbor vehicles,
current y coordinates of the target vehicle and the one or more neighbor vehicles,
previous x coordinates of the target vehicle and the one or more neighbor vehicles,
previous y coordinates of the target vehicle and the one or more neighbor vehicles,
x-axis velocity values of the target vehicle and the one or more neighbor vehicles,
y-axis velocity values of the target vehicle and the one or more neighbor vehicles,
angular velocity values of the target vehicle and the one or more neighbor vehicles,
acceleration values of the target vehicle and the one or more neighbor vehicles,
blinker states of the target vehicle and the one or more neighbor vehicles,
a time offset between current and previous positions of the target vehicle and the one or more neighbor vehicles,
flags indicating a presence of the target vehicle and the one or more neighbor vehicles, or
any combination thereof.
19 . The method of claim 16 , wherein the one or more map tensors represent map context information around the target vehicle.
20 . The method of claim 19 , wherein the map context information comprises:
map lane center points for a current lane of the target vehicle,
map lane center points for a lane to the left of the target vehicle,
map lane center points for a lane to the right of the target vehicle, or
any combination thereof.
21 . The method of claim 16 , wherein an encoder side of the encoder-decoder machine learning model comprises:
an agent encoding module,
a map encoding module, and
an aggregation module.
22 . The method of claim 21 , wherein the agent encoding module comprises a self-attention module.
23 . The method of claim 21 , wherein the aggregation module aggregates results of the agent encoding module and the map encoding module.
24 . The method of claim 21 , wherein the aggregation module comprises a map attention module and a cross attention module.
25 . The method of claim 16 , wherein the decoder side of the encoder-decoder machine learning model comprises:
one or more MLP layers that output the lane change classification label.
26 . The method of claim 16 , wherein the predicted lane change intention further includes one or more predicted lane change trajectories associated with the lane change classification.
27 . An ego vehicle, comprising:
means for applying an encoder-decoder machine learning model to one or more agent tensors and one or more map tensors associated with a target vehicle to obtain a lane change classification label representing a predicted lane change intention of the target vehicle on a multi-lane highway, wherein the predicted lane change intention includes at least a lane change classification, and wherein a decoder side of the encoder-decoder machine learning model outputs the lane change classification label; and
means for performing a driving maneuver based on the predicted lane change intention.
28 . A non-transitory computer-readable medium storing computer-executable instructions that, when executed by an ego vehicle, cause the ego vehicle to:
apply an encoder-decoder machine learning model to one or more agent tensors and one or more map tensors associated with a target vehicle to obtain a lane change classification label representing a predicted lane change intention of the target vehicle on a multi-lane highway, wherein the predicted lane change intention includes at least a lane change classification, and wherein a decoder side of the encoder-decoder machine learning model outputs the lane change classification label; and
perform a driving maneuver based on the predicted lane change intention.