Vulnerable road user detection
Methods, systems, and computer programs are presented to detect potential collisions of vehicles with a Vulnerable Road User (VRU). The provided solution includes a system designed to detect and warn against potential collisions with vulnerable road users (VRUs) using a combination of hardware and software components. The system employs one or more cameras installed on a vehicle to monitor the surrounding environment. These cameras capture image frames, which are processed by an onboard model to estimate potential collisions through trajectory analysis. In one aspect, the trajectory estimation includes using a bird's eye view transformation, which provides a top-down perspective of the scene to aid in trajectory estimation. Camera parameters are personalized per installation by determining a transformation matrix used to derive the bird's eye view.
1 . A computer-implemented method comprising:
identifying a location of a vulnerable road user (VRU) in a sequence of images taken by a camera on a vehicle;
estimating a trajectory of the VRU based on the location of the VRU in the sequence of images, wherein estimating the trajectory of the VRU further comprises:
transforming each image from the sequence of images to a bird's eye view using a homography matrix associated with the camera in the vehicle; and
estimating the trajectory of the VRU based on a sequence of bird's eye views;
estimating a trajectory of the vehicle based on the sequence of images; and
determining if a collision is predicted from the vehicle to the VRU based on the trajectory of the VRU and the trajectory of the vehicle.
2 . The method as recited in claim 1 , wherein identifying the location of the VRU further comprises:
using a machine learning (ML) model trained to detect the VRU based on an input image, the ML model generating as output a bounding box of the VRU in the input image.
3 . The method as recited in claim 2 , wherein the ML model assigns confidence scores to each detected VRU indicating a likelihood of accurate detection.
4 . The method as recited in claim 1 , wherein estimating the trajectory of the vehicle further comprises:
determining a direction of the vehicle associated with a most recent image.
5 . The method of claim 1 , further comprising:
calculating the homography matrix based on correspondences between known real-world points and positions of the known real-world points in the images, wherein the homography matrix is a transformation matrix that maps points from an image with an original perspective to a top-down orthogonal view.
6 . The method of claim 5 , wherein the calculation of the homography matrix for the camera installed in the vehicle is performed on a remote server, which is downloaded from the remote server to an asset monitoring system in the vehicle.
7 . The method of claim 1 , further comprising:
extracting, by an ML model combined with a least squares solver, camera parameters of the camera installed in the vehicle based on frames captured by the camera, the camera parameters comprising focal length, roll, pitch, and yaw; and
customizing the homography matrix for the camera installed in the vehicle based on the camera parameters.
8 . The method as recited in claim 1 , further comprising:
overlaying segmentation masks onto a global depth mask to determine object depth within the image.
9 . The method as recited in claim 1 , wherein determining if the collision is predicted further comprises:
calculating a severity score for a probability of collision based on object depth, position, and motion state.
10 . The method as recited in claim 1 , wherein determining if the collision is predicted further comprises:
detecting a possible collision in a predetermined minimum number of consecutive images in the sequence of images; and
generating an alert based on the possible collision being detected in the predetermined minimum number of consecutive images.
11 . The method as recited in claim 1 , wherein tracked VRUs comprise pedestrians, cyclists, motorcyclists, users of mobility devices, and animal-related users.
12 . The method as recited in claim 1 , wherein the sequence of images is sampled from a video at a predefined sampling rate.
13 . A system comprising:
a memory comprising instructions; and
one or more computer processors, the instructions, when executed by the one or more computer processors, causing the system to perform operations comprising:
identifying a location of a vulnerable road user (VRU) in a sequence of images taken by a camera on a vehicle;
estimating a trajectory of the VRU based on the location of the VRU in the sequence of images, wherein estimating the trajectory of the VRU further comprises:
transforming each image from the sequence of images to a bird's eye view using a homography matrix associated with the camera in the vehicle; and
estimating the trajectory of the VRU based on a sequence of bird's eye views;
estimating a trajectory of the vehicle based on the sequence of images; and
determining if a collision is predicted from the vehicle to the VRU based on the trajectory of the VRU and the trajectory of the vehicle.
14 . The system as recited in claim 13 , wherein identifying the location of the VRU further comprises:
using a machine learning (ML) model trained to detect the VRU based on an input image, the ML model generating as output a bounding box of the VRU in the input image.
15 . The system as recited in claim 14 , wherein the ML model assigns confidence scores to each detected VRU indicating a likelihood of accurate detection.
16 . The system as recited in claim 13 , wherein estimating the trajectory of the vehicle further comprises:
determining a direction of the vehicle associated with a most recent image.
17 . A non-transitory machine-readable storage medium including instructions that, when executed by a machine, cause the machine to perform operations comprising:
identifying a location of a vulnerable road user (VRU) in a sequence of images taken by a camera on a vehicle;
estimating a trajectory of the VRU based on the location of the VRU in the sequence of images, wherein estimating the trajectory of the VRU further comprises:
transforming each image from the sequence of images to a bird's eye view using a homography matrix associated with the camera in the vehicle; and
estimating the trajectory of the VRU based on a sequence of bird's eye views;
estimating a trajectory of the vehicle based on the sequence of images; and
determining if a collision is predicted from the vehicle to the VRU based on the trajectory of the VRU and the trajectory of the vehicle.
18 . The non-transitory machine-readable storage medium as recited in claim 17 , wherein identifying the location of the VRU further comprises:
using a machine learning (ML) model trained to detect the VRU based on an input image, the ML model generating as output a bounding box of the VRU in the input image.
19 . The non-transitory machine-readable storage medium as recited in claim 18 , wherein the ML model assigns confidence scores to each detected VRU indicating a likelihood of accurate detection.
20 . The non-transitory machine-readable storage medium as recited in claim 17 , wherein estimating the trajectory of the vehicle further comprises:
determining a direction of the vehicle associated with a most recent image.