Systems and methods for training a model for determining vehicle following distance
Systems, methods, models, and training data for models are discussed, for determining vehicle positioning, and in particular identifying tailgating. Simulated training images showing vehicles following other vehicles, under various conditions, are generated using a virtual environment. Models are trained to determine following distance between two vehicles. Trained models are used to in detection of tailgating, based on determined distance between two vehicles. Results of tailgating are output to warn a driver, or to provide a report on driver behavior.
1 . A system for training a model for determining a distance between two vehicles, the system comprising:
at least one processor;
at least one non-transitory processor-readable storage medium communicatively coupled to the at least one processor and storing processor-executable instructions which when executed by the at least one processor cause the system to:
render image data based on simulation in a virtual environment, the image data including at least:
a first set of images, each image in the first set of images including a representation of a respective first virtual vehicle from a perspective of a respective second virtual vehicle behind the respective first virtual vehicle, and each image in the first set of images associated with; a respective distance label indicating a distance between the respective first virtual vehicle and the respective second virtual vehicle and a respective vehicle presence label indicating that the respective first virtual vehicle is within a vehicle presence threshold distance of the respective second virtual vehicle; and
a second set of images, each image in the second set of images including a representation from a perspective of a respective third virtual vehicle which is not in a following situation, and each image in the second set of images associated with: a respective distance label that is a null value or infinity label indicative of a non-following situation, and a respective vehicle presence label indicating that the respective third virtual vehicle is not within the vehicle presence threshold distance of another vehicle;
generating respective vehicle presence labels and respective distance labels at the time of simulation in the virtual environment based on ground-truth relative positions of the respective first virtual vehicle and the respective second virtual vehicle, or the respective third virtual vehicle;
evaluate a following distance loss function for at least one image in the first set of images and at least one image in the second set of images, the following distance loss function including: a first term comprising the vehicle presence label multiplied by a difference between a distance indicated in a respective distance label and a determined distance between two vehicles by the model for each respective image, and a second term representing a difference between the vehicle presence label and a determined vehicle presence for each respective image; and
train the model by minimizing the following distance loss function over the first set of images and the second set of images.
2 . The system of claim 1 , wherein the processor-executable instructions further cause the system to:
determine whether auxiliary criteria are satisfied over the first set of images; and
further evaluate the following distance loss function for at least one image in the first set of images, if the auxiliary criteria are not satisfied.
3 . The system of claim 2 , wherein the auxiliary criteria require that the following distance loss function be within a maximum loss threshold for each image in the first set of images.
4 . The system of claim 2 , wherein the auxiliary criteria require that the following distance loss function be within a maximum loss threshold for a defined quantity of images in the first set of images, where the defined quantity of images is smaller than a total quantity of images in the first set of images.
5 . The system of claim 2 , wherein the auxiliary criteria require that the following distance loss function be evaluated for each image in the first set of images.
6 . The system of claim 2 , wherein the auxiliary criteria require that the following distance loss function be evaluated for a defined quantity of images in the first set of images, where the defined quantity of images is smaller than a total quantity of images in the first set of images.
7 . The system of claim 1 , wherein:
the processor-executable instructions which cause the at least one processor to generate the respective vehicle presence labels cause the at least one processor to generate the vehicle presence labels further based on vehicle lane of travel; and
for each image of the first set of images, the vehicle presence label indicates presence when both: (i) the respective first virtual vehicle is within a vehicle presence threshold distance of the respective second virtual vehicle; and (ii) the respective first virtual vehicle is travelling in a same lane of travel as the respective second virtual vehicle.
8 . The system of claim 1 , wherein the processor-executable instructions cause the at least one processor to automatically determine random positions within the virtual environment for each respective first virtual vehicle and each respective second virtual vehicle, where the random positions for are constrained such that each respective first virtual vehicle is in a same lane of travel as a corresponding second virtual vehicle.
9 . The system of claim 1 , wherein the determined vehicle presence for each respective image in the second term is determined based on at least one of:
detected lane boundaries in the respective image; and
calibrated horizontal-distance criteria between respective left-boundaries and right-boundaries of a bounding box of a respective first vehicle and image edges of the respective image.
10 . The system of claim 1 , wherein the determined vehicle presence for each respective image in the second term is a confidence value between 0 and 1 indicating a probability that the first vehicle is within the vehicle presence threshold.
11 . A method for training a model for determining a distance between two vehicles, the method comprising:
rendering image data based on simulation in a virtual environment, the image data including at least:
a first set of images, each image in the first set of images including a representation of a respective first virtual vehicle from a perspective of a respective second virtual vehicle behind the respective first virtual vehicle, and each image in the first set of images associated with a distance label indicating a distance between the respective first virtual vehicle and the respective second virtual vehicle and a respective vehicle presence label indicating that the respective first virtual vehicle is within a vehicle presence threshold distance of the respective second virtual vehicle; and
a second set of images, each image in the second set of images including a representation from a perspective of a respective third virtual vehicle which is not in a following situation, and each image in the second set of images associated with: a respective distance label that is a null value or infinity label indicative of a non-following situation, and a respective vehicle presence label indicating that the respective third virtual vehicle is not within the vehicle presence threshold distance of another vehicle;
generating respective vehicle presence labels and respective distance labels at the time of simulation in the virtual environment based on ground-truth relative position of the respective first virtual vehicle and the respective second virtual vehicle, or the respective third virtual vehicle;
evaluating a following distance loss function for at least one image in the first set of images and at least one image in the second set of images, the following distance loss function including; a first term comprising the vehicle presence label multiplied by a difference between a distance indicated in a respective distance label and a determined distance between two vehicles by the model for each respective image, and a second term representing a difference between the vehicle presence label and a determined vehicle presence for each respective image; and
training the model by minimizing the following distance loss function over the first set of images and the second set of images.
12 . The method of claim 11 , further comprising:
determining whether auxiliary criteria are satisfied over the first set of images; and
further evaluating the following distance loss function for at least one image in the first set of images, if the auxiliary criteria are not satisfied.
13 . The method of claim 12 , wherein the auxiliary criteria require that the following distance loss function be within a maximum loss threshold for each image in the first set of images.
14 . The method of claim 12 , wherein the auxiliary criteria require that the following distance loss function be within a maximum loss threshold for a defined quantity of images in the first set of images, where the defined quantity of images is smaller than a total quantity of images in the first set of images.
15 . The method of claim 12 , wherein the auxiliary criteria require that the following distance loss function be evaluated for each image in the first set of images.
16 . The method of claim 12 , wherein the auxiliary criteria require that the following distance loss function be evaluated for a defined quantity of images in the first set of images, where the defined quantity of images is smaller than a total quantity of images in the first set of images.
17 . The method of claim 11 , wherein:
generating the respective vehicle presence labels comprises generating the vehicle presence labels further based on vehicle lane of travel; and
for each image of the first set of images, the vehicle presence label indicates presence when both: (i) the respective first virtual vehicle is within a vehicle presence threshold distance of the respective second virtual vehicle; and (ii) the respective first virtual vehicle is travelling in a same lane of travel as the respective second virtual vehicle.
18 . The method of claim 11 , further comprising automatically determining random positions within the virtual environment for each respective first virtual vehicle and each respective second virtual vehicle, where the random positions for are constrained such that each respective first virtual vehicle is in a same lane of travel as a corresponding second virtual vehicle.
19 . The method of claim 11 , wherein the determined vehicle presence for each respective image in the second term is determined based on at least one of:
detected lane boundaries in the respective image; and
calibrated horizontal-distance criteria between respective left-boundaries and right-boundaries of a bounding box of a respective first virtual vehicle and image edges of the respective image.
20 . The method of claim 11 , wherein the determined vehicle presence for each respective image in the second term is a confidence value between 0 and 1 indicating a probability that the respective first virtual vehicle is within the vehicle presence threshold.