Search

EP-4283252-B1 - METHOD AND SYSTEM FOR CLASSIFICATION OF AN OBJECT IN A POINT CLOUD DATA SET

EP4283252B1EP 4283252 B1EP4283252 B1EP 4283252B1EP-4283252-B1

Inventors

  • CROUCH, STEPHEN C
  • REIBEL, RANDY R.
  • KAYLOR, BRANT

Dates

Publication Date
20260506
Application Date
20171121

Claims (12)

  1. A light detection and ranging, LIDAR, sensor system (200) for a vehicle, comprising: a sensor configured to: transmit a transmit beam (126; 205), the transmit beam having at least one of a frequency or a phase that is modulated over time; receive a return beam (291; 391) from reflection or scattering of the transmit beam by an object (390); and output a data signal comprising a plurality of point cloud data points (400; 600) representing the object; and one or more processors (1102; 1203) configured to: determine at least one feature variable (α; β; ρ; Ψ) based on spatial relationships between a number of points (605) of the plurality of point cloud data points and a reference point (601) of the plurality of point cloud data points, wherein the number of points (605) is defined within a translational and rotational invariant coordinate system defined by a surface normal (602) at the reference point (601) and a plane (655) that is orthogonal to the surface normal (602) and tangent to a surface (656) of the object at the reference point (601); determine a first classification statistic (680) based on values of the at least one feature variable, wherein the first classification statistic is a spin image; determine a second classification statistic based on a covariance matrix based on the values of the at least one feature variable; and determine an object class corresponding to the object based on the first classification statistic and the second classification statistic, characterized in that determining the object class comprises: determining a closest match between the first classification statistic and a set of first classification statistics for a corresponding set of N classes to estimate that the object generating the point cloud (600) is in a first class of the set of N classes; determining a closest match between the second classification statistic and a set of second classification statistics for the corresponding set of N classes to estimate that the object generating the point cloud (600) is in a second class of the set of N classes; and assigning the object to the first class responsive to the first class being the same as the second class.
  2. The LIDAR sensor system of claim 1, further comprising: a laser source (212) configured to generate a carrier wave (201) modulated in at least one of the frequency or the phase; an optic configured to output the carrier wave as the transmit beam; one or more detectors (230; 330) configured to output an electrical signal based on the return beam.
  3. The LIDAR sensor system of claim 1, wherein the sensor comprises: a laser source (212) configured to output a carrier wave (201); a modulator (214; 310) configured to modulate the carrier wave to provide the carrier wave as the transmit beam; and one or more scanning optics (320) configured to transmit the transmit beam.
  4. The LIDAR sensor system of claim 3, wherein the modulator is configured to modulate the at least one of the phase or the frequency by applying modulation to a current driving the laser source.
  5. The LIDAR sensor system of claim 3, wherein the modulator is an electro-optic modulator or an acoustic-optic modulator.
  6. The LIDAR sensor system of claim 1, wherein the transmit beam is a chirp signal having a bandwidth and a duration, wherein a frequency of the chirp signal increases over the duration of the chirp signal.
  7. The LIDAR sensor system of claim 1, wherein the sensor is further configured to output the data signal as a three-dimensional, 3D, point cloud comprising the plurality of point cloud data points.
  8. The LIDAR sensor system of claim 1, wherein the one or more processors are configured to determine the object class from a predetermined number of object classes.
  9. The LIDAR sensor system of claim 1, wherein the one or more processors are configured to determine the object class from a vehicle class and a roadside structure class.
  10. The LIDAR sensor system of claim 1, wherein the one or more processors are configured to: determine the first classification statistic, which is the spin image, based on a histogram of at least a subset of the plurality of point cloud data points in each of a plurality of bins corresponding to a range of values of the at least one feature variable.
  11. The LIDAR sensor system of any one of claims 1 to 10, wherein the one or more processors are further configured to avoid collision with the object based on the determined class.
  12. A method, comprising: transmitting a transmit beam (126; 205) having at least one of a frequency or a phase that is modulated over time; receiving a return beam (291; 391) from reflection or scattering of the transmit beam by an object (390); outputting a data signal comprising a plurality of point cloud data points (400; 600) representing the object; determining at least one feature variable (α; β; ρ; ψ) based on spatial relationships between a number of points (605) of the plurality of point cloud data points and a reference point (601) of the plurality of point cloud data points, wherein the number of points (605) is defined within a translational and rotational invariant coordinate system defined by a surface normal (602) at the reference point (601) and a plane (655) that is orthogonal to the surface normal (602) and tangent to a surface (656) of the object at the reference point (601); determining a first classification statistic (680) based on values of the at least one feature variable, wherein the first classification statistic is a spin image; determining a second classification statistic based on a covariance matrix based on the values of the at least one feature variable; and determining an object class corresponding to the object based on the first classification statistic and the second classification statistic, characterized in that determining the object class comprises: determining a closest match between the first classification statistic and a set of first classification statistics for a corresponding set of N classes to estimate that the object generating the point cloud (600) is in a first class of the set of N classes; determining a closest match between the second classification statistic and a set of second classification statistics for the corresponding set of N classes to estimate that the object generating the point cloud (600) is in a second class of the set of N classes; and assigning the object to the first class responsive to the first class being the same as the second class.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application claims benefit of Provisional Appln. 62/427,573, filed November 29, 2016. BACKGROUND Optical detection of range, often referenced by a mnemonic, LIDAR, for light detection and ranging, is used for a variety of applications, from altimetry, to imaging, to collision avoidance. LIDAR provides finer scale range resolution with smaller beam sizes than conventional microwave ranging systems, such as radio-wave detection and ranging (RADAR). Optical detection of range can be accomplished with several different techniques, including direct ranging based on round trip travel time of an optical pulse to a target, and chirped detection based on a frequency difference between a transmitted chirped optical signal and a returned signal scattered from a target, and phase encoded detection based on a sequence of single frequency phase changes that are distinguishable from natural signals. To achieve acceptable range accuracy and detection sensitivity, direct long range LIDAR systems use short pulse lasers with low pulse repetition rate and extremely high pulse peak power. The high pulse power can lead to rapid degradation of optical components. Chirped LIDAR systems use long optical pulses with relatively low peak optical power. In this configuration, the range accuracy depends on the chirp bandwidth rather than the pulse duration, and therefore excellent range accuracy can still be obtained. Useful optical chirp bandwidths have been achieved using wideband radio frequency (RF) electrical signals to modulate an optical carrier. Recent advances in chirped LIDAR include using the same modulated optical carrier as a reference signal that is combined with the returned signal at an optical detector to produce in the resulting electrical signal a relatively low beat frequency that is proportional to the difference in frequencies between the references and returned optical signals. This kind of beat frequency detection of frequency differences at a detector is called heterodyne detection. It has several advantages known in the art, such as the advantage of using RF components of ready and inexpensive availability. Recent work described in U.S. Patent Number 7,742,152, shows a novel simpler arrangement of optical components that uses, as the reference optical signal, an optical signal split from the transmitted optical signal. This arrangement is called homodyne detection in that patent. LIDAR detection with phase encoded microwave signals modulated onto an optical carrier have been used as well. This technique relies on correlating a sequence of phases (or phase changes) of a particular frequency in a return signal with that in the transmitted signal. A time delay associated with a peak in correlation is related to range by the speed of light in the medium. Advantages of this technique include the need for fewer components, and the use of mass produced hardware components developed for phase encoded microwave and optical communications. The data returned by these LIDAR systems is often represented as a point cloud. A point cloud is a set of data points in some coordinate system. In a three dimensional coordinate system, these points are usually defined by X, Y and Z coordinates, and often are intended to represent the external surface of an object. 3D point clouds can be generated by 3D scanners, such as LIDAR systems including chirped LIDAR and phase coded LIDAR, among other types of scanners. The paper "A multiple classifier system for classification of LIDAR remote sensing data using multi-class SVM" (Samadzadegan et al., 2010) applies a multi-class SVM based classifier fusion technique on LIDAR data. In a nutshell, different subsets of features, such as NDI, Entropy, Mean, Standard Deviation, etc. are used to train and produce a plurality of classifiers. The final classification decision is obtained by fusing the results of SVM classifiers using weighted majority voting. Moreover, the results obtained when using individual SVM classifiers are then compared to the results of fusion in terms of the accuracy achieved by said individual SVM classifier. Next, overall accuracy of each classifier is used as its weight in weighted Voting method. Thus, the method aims at using a plurality of classifiers (8 in the example of that paper) and then weighting their contribution depending on the prediction accuracy of each individual classifier. SUMMARY The current inventors have recognized circumstances and applications in which automatic classification of objects represented by 3D point clouds is challenging in real time, particularly for objects located at long range. Techniques are provided for such automatic classification of objects. The invention is defined by the appended claims. According to a first aspect of the present invention, there is provided a light detection and ranging, LIDAR, sensor system for a vehicle as set out in claim 1. According to a second aspect of the pre