EP-4738292-A1 - METHOD AND APPARATUS FOR DETECTING STATIONARY OBJECT, AND DEVICE, STORAGE MEDIUM AND VEHICLE

EP4738292A1EP 4738292 A1EP4738292 A1EP 4738292A1EP-4738292-A1

Abstract

The present application relates to the technical field of object detection. Disclosed are a method and apparatus for detecting a stationary object, and a device, a storage medium and a vehicle. The method comprises: acquiring the current image frame collected by a camera, and identifying a camera observation object from the current image frame; associating the camera observation object with a current object detection box, and performing state prediction and update on the associated object detection box on the basis of a relative motion parameter of the camera observation object, wherein the current object detection box comprises a detection box created on the basis of historical observed objects; determining whether the state of the updated object detection box has converged; and when the state of the updated object detection box has converged, determining whether the updated object detection box is stationary. In the present application, an observation object is associated with an object detection box and a state is updated, such that whether the object detection box is stationary can be accurately determined by using the object detection box with a converged state.

Inventors

DENG, Huan
YU, FENGNING
DIAO, Pengfei
LU, Di
CHEN, QI
XIAO, Yao
YANG, Fanfan
SONG, Jiarong
LI, Tingfei
MU, Lewen

Assignees

Guangzhou Xiaopeng Autopilot Technology Co., Ltd.

Dates

Publication Date: 20260506
Application Date: 20240815

Claims (14)

A method for detecting a stationary target, characterized in that the method comprises: obtaining a current frame image captured by a camera, and identifying a camera-observed target in the current frame image; associating the camera-observed target with a current detection bounding box, and performing state prediction and update on the associated detection bounding box according to a relative motion parameter of the camera-observed target, wherein the current detection bounding box comprises a detection bounding box created based on a historically observed target; determining whether a state of the updated detection bounding box has converged; and when the state of the updated detection bounding box has converged, determining whether the updated detection bounding box is stationary.
The method according to claim 1, wherein the associating the camera-observed target with a current detection bounding box comprises: associating a first camera-observed target with the current detection bounding box, wherein the first camera-observed target is a camera-observed target corresponding to a most critical dynamic object; thereafter, associating a second camera-observed target with a detection bounding box currently not associated with the first camera-observed target, wherein the second camera-observed target is a camera-observed target other than the first camera-observed target; and creating a new detection bounding box for a second camera-observed target that fails to be associated with a detection bounding box.
The method according to claim 2, wherein the associating the camera-observed target with a current detection bounding box further comprises: in a case where the first camera-observed target fails to be associated with a detection bounding box, after creating a new detection bounding box for a second camera-observed target that fails to be associated with a detection bounding box, re-associating the first camera-observed target with a detection bounding box currently not associated with a camera-observed target; and in a case where the first camera-observed target fails to be associated with a detection bounding box again, creating a new detection bounding box for the first camera-observed target.
The method according to claim 1, wherein the determining whether a state of the updated detection bounding box has converged comprises: determining a first number of associations of the updated detection bounding box, wherein the first number of associations is the number of times that the updated detection bounding box is associated with a camera-observed target corresponding to a most critical dynamic object; and in a case where the first number of associations is greater than a first threshold, determining that the state of the updated detection bounding box has converged, and/or determining a second number of associations of the updated detection bounding box, wherein the second number of associations is the number of times that the updated detection bounding box is associated with a camera-observed target; in a case where the second number of associations is greater than a second threshold, determining a covariance matrix between a relative speed and a relative depth corresponding to the updated detection bounding box; and in a case where the covariance matrix satisfies a convergence condition, determining that the state of the updated detection bounding box has converged.
The method according to claim 4, wherein the covariance matrix is: P depth speed = P dd P ds P sd P ss , where depth denotes a relative depth observation, speed denotes a relative speed observation, P ( depth,speed ) denotes the covariance matrix between the relative speed and the relative depth, P dd denotes a variance of the relative depth observation, P ss denotes a variance of the relative speed observation, P ds denotes a covariance between the relative depth observation and the relative speed observation, P sd denotes a covariance between the relative speed observation and the relative depth observation; and the convergence condition is: P dd − P ds ∗ P sd / P ss < k 1 ∗ depth ′ 2 P ss − P sd ∗ P ds / P dd < k 2 2 , where depth' denotes a relative depth of the detection bounding box, k 1 and k 2 are both preset threshold parameters.
The method according to claim 1, wherein the determining whether the updated detection bounding box is stationary when the state of the updated detection bounding box has converged comprises: in a case where the state of the updated detection bounding box has converged, determining that the updated detection bounding box is stationary if a relative speed of the updated detection bounding box is less than a current speed threshold, or in a case where the state of the updated detection bounding box has converged, determining a normal distribution cumulative probability corresponding to an ego vehicle speed by using a negative value of the ego vehicle speed as a variable and using a relative speed of the updated detection bounding box as an expectation; and in a case where the normal distribution cumulative probability is greater than a preset probability value, determining that the updated detection bounding box is stationary.
The method according to claim 6, wherein the normal distribution cumulative probability is: F x μ σ = 1 σ 2 π ∫ − ∞ x + Δ v exp − x − μ 2 2 σ 2 dx , where x denotes a negative value of the ego vehicle speed, Δ v denotes a preset speed difference, µ denotes a relative speed expectation, σ denotes a relative speed standard deviation.
The method according to claim 1, wherein the method further comprises: determining a radar-observed target currently identified by a radar, and identifying a stationary radar-observed target; and associating the stationary radar-observed target with a current detection bounding box, and performing state prediction and update on the associated detection bounding box according to a relative motion parameter of the stationary radar-observed target.
The method according to claim 8, wherein: the determining whether a state of the updated detection bounding box has converged comprises: determining a third number of associations of the updated detection bounding box, wherein the third number of associations is the number of times that the updated detection bounding box is associated with the stationary radar-observed target; and in a case where the third number of associations is greater than a third threshold, determining that the state of the updated detection bounding box has converged; and the determining whether the updated detection bounding box is stationary in a case where the state of the updated detection bounding box has converged comprises: in a case where the third number of associations that the updated detection bounding box is associated with the stationary radar-observed target is greater than the third threshold, determining that the updated detection bounding box is stationary.
The method according to claim 1, wherein the method further comprises: deleting a detection bounding box meeting a failure condition, wherein the failure condition comprises: a duration of a detection bounding box not being associated exceeding a first preset duration; or a target type of a detection bounding box being inconsistent with a target type of an associated observed target and a duration of being inconsistent exceeding a second preset duration.
An apparatus for detecting a stationary target, characterized in that , the apparatus comprises: a target identification module configured to obtain a current frame image captured by a camera and identify a camera-observed target in the current frame image; an association and update module configured to associate the camera-observed target with a current detection bounding box and perform state prediction and update on the associated detection bounding box according to a relative motion parameter of the camera-observed target, wherein the current detection bounding box comprises a bounding box created based on a historically observed target; a convergence determination module configured to determine whether a state of the updated detection bounding box has converged; and a detection module configured to determine whether the updated detection bounding box is stationary in a case where the state of the updated detection bounding box has converged.
A computer device, characterized in that , the device comprises: a memory and a processor in communication with each other, the memory having stored thereon computer instructions which, when executed by the processor, cause the processor to execute the method according to any of claims 1-10.
A computer-readable storage medium having stored thereon computer instructions which cause a computer to perform the method according to any of claims 1-10.
A vehicle, comprising a vehicle controller configured to perform the method according to any of claims 1-10.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS The present application claims priority to Chinese Patent Application No. 202410972892.5, filed on July 18, 2024 and entitled "Method for detecting a stationary target, apparatus, device, storage medium and vehicle therefor", which is incorporated herein by reference in its entirety. TECHNICAL FIELD The present application relates to the field of target detection technology, and in particular, to a method for detecting a stationary target, apparatus, device, storage medium and vehicle therefor. BACKGROUND Intelligent driving vehicles are equipped with various sensors such as cameras and radars. Based on data collected by the sensors, objects around the vehicle can be detected to avoid collisions. Stationary objects have a significant impact on vehicle movement and it needs to accurately detect those around the vehicle. When detecting a stationary target, especially one at a long distance (for example, a distance of more than 100m), speed and distance measuring accuracy of cameras are not high, while radars often produce false detections of stationary targets. Some solutions integrate features of images from cameras and radar features, but data fusion requires a large amount of computation and has low processing efficiency. SUMMARY In view of this, the present application provides a method, apparatus, device, storage medium and vehicle for detecting a stationary target in order to address the issue of inaccurate detection of the stationary target. According to a first aspect of the present application, a method for detecting a stationary target is provided. The method comprises: obtaining a current frame image captured by a camera, and identifying a camera-observed target in the current frame image;associating the camera-observed target with a current detection bounding box, and performing state prediction and update on the associated detection bounding box according to a relative motion parameter of the camera-observed target, wherein the current detection bounding box comprises a detection bounding box created based on a historically observed target;determining whether a state of the updated detection bounding box has converged; andwhen the state of the updated detection bounding box has converged, determining whether the updated detection bounding box is stationary. In some optional implementations, the associating the camera-observed target with a current detection bounding box comprises: associating a first camera-observed target with the current detection bounding box, wherein the first camera-observed target is a camera-observed target corresponding to a most critical dynamic object;thereafter, associating a second camera-observed target with a detection bounding box currently not associated with the first camera-observed target, wherein the second camera-observed target is a camera-observed target other than the first camera-observed target; andcreating a new detection bounding box for a second camera-observed target that fails to be associated with a detection bounding box. In some optional implementations, the associating the camera-observed target with a current detection bounding box further comprises: in a case where the first camera-observed target fails to be associated with a detection bounding box, after creating a new detection bounding box for a second camera-observed target that fails to be associated with a detection bounding box, re-associating the first camera-observed target with a detection bounding box currently not associated with a camera-observed target; andin a case where the first camera-observed target fails to be associated with a detection bounding box again, creating a new detection bounding box for the first camera-observed target. In some optional implementations, the determining whether a state of the updated detection bounding box has converged comprises: determining a first number of associations of the updated detection bounding box, wherein the first number of associations is the number of times that the updated detection bounding box is associated with a camera-observed target corresponding to a most critical dynamic object; andin a case where the first number of associations is greater than a first threshold, determining that the state of the updated detection bounding box has converged, and/ordetermining a second number of associations of the updated detection bounding box, wherein the second number of associations is the number of times that the updated detection bounding box is associated with a camera-observed target;in a case where the second number of associations is greater than a second threshold, determining a covariance matrix between a relative speed and a relative depth corresponding to the updated detection bounding box; andin a case where the covariance matrix satisfies a convergence condition, determining that the state of the updated detection bounding box has converged. In some optional implementations, the c