Patent Application
20200057160
This present disclosure is directed generally to electronic signal processing, and more specifically, to signal processing associated components, systems and techniques in light detection and ranging (LIDAR) applications.
With their ever increasing performance and lowering cost, unmanned movable objects, such as unmanned robotics, are now extensively used in many fields. Representative missions include real estate photography, inspection of buildings and other structures, fire and safety missions, border patrols, and product delivery, among others. For obstacle detection as well as for other functionalities, it is beneficial for the unmanned vehicles to be equipped with obstacle detection and surrounding environment scanning devices. Light detection and ranging (LIDAR, also known as “light radar”) is a reliable and stable detection technology. However, traditional LIDAR devices are typically expensive because they use multi-channel, high-density, and high-speed emitters and sensors, making most traditional LIDAR devices unfit for low cost unmanned vehicle applications.
Accordingly, there remains a need for improved techniques and systems for implementing LIDAR scanning modules, for example, such as those carried by unmanned vehicles and other objects.
This patent document relates to techniques, systems, and devices for conducting object tracking by an unmanned vehicle using multiple low-cost LIDAR emitter and sensor pairs.
In one exemplary aspect, a light detection and ranging (LIDAR) based object tracking system is disclosed. The system includes a plurality of light emitter and sensor pairs. Each pair of the plurality of light emitter and sensor pairs is operable to obtain data indicative of actual locations of surrounding objects. The data is grouped into a plurality of groups by a segmentation module, each group corresponding to one of the surrounding objects. The system also includes an object tracker configured to (1) build a plurality of models of target objects based on the plurality of groups, (2) compute a motion estimation for each of the target objects, and (3) feed a subset of data back to the segmentation module for further grouping based on a determination by the object tracker that the subset of data fails to map to a corresponding target object in the model.
In another exemplary aspect, a microcontroller system for controlling an unmanned movable object is disclosed. The system includes a processor configured to implement a method of tracking objects in real-time or near real-time. The method includes receiving data indicative of actual locations of surrounding objects. The actual locations are grouped into a plurality of groups by a segmentation module, and each group of the plurality of groups corresponds to one of the surrounding objects. The method also includes obtaining a plurality of models of target objects based on the plurality of groups, estimating a motion matrix for each of the target objects, updating the model using the motion matrix for each of the target objects, and optimizing the model by modifying the model for each of the target objects to remove or reduce a physical distortion of the model for the target object.
In yet another exemplary aspect, an unmanned device is disclosed. The unmanned device includes light detection and ranging (LIDAR) based object tracking system as described above, a controller operable to generate control signals to direct motion of the vehicle in response to output from the real-time object tracking system, and an engine operable to maneuver the vehicle in response to control signals from the controller.
The above and other aspects and their implementations are described in greater detail in the drawings, the description and the claims.
With the ever increasing use of unmanned movable objects, such as unmanned vehicles, it is important for them to be able to independently detect obstacles and to automatically engage in obstacle avoidance maneuvers. Light detection and ranging (LIDAR) is a reliable and stable detection technology because LIDAR can remain functional under nearly all weather conditions. Moreover, unlike traditional image sensors (e.g., cameras) that can only sense the surroundings in two dimensions, LIDAR can obtain three-dimensional information by detecting the depth. However, traditional LIDAR systems are typically expensive because they rely on multi-channel, high-speed, high-density LIDAR emitters and sensors. The cost of such LIDARs, together with the cost of having sufficient processing power to process the dense data, makes the price of traditional LIDAR systems formidable. This patent document describes techniques and methods for utilizing multiple low-cost single-channel linear LIDAR emitter and sensor pairs to achieve multi-object tracking by unmanned vehicles. The disclosed techniques are capable of achieving multi-object tracking with a much lower data density (e.g., around 1/10 of the data density in traditional approaches) while maintaining similar precision and robustness for object tracking.
In the following description, the example of a unmanned vehicle is used, for illustrative purposes only, to explain various techniques that can be implemented using a LIDAR object tracking system that is more cost-effective than the traditional LIDARs. For example, even though one or more figures introduced in connection with the techniques illustrate a unmanned car, in other embodiments, the techniques are applicable in a similar manner to other type of movable objects including, but not limited to, an unmanned aviation vehicle, a hand-held device, or a robot. In another example, even though the techniques are particularly applicable to laser beams produced by laser diodes in a LIDAR system, the scanning results from other types of object range sensor, such as a time-of-flight camera, can also be applicable.
In the following, numerous specific details are set forth to provide a thorough understanding of the presently disclosed technology. In some instances, well-known features are not described in detail to avoid unnecessarily obscuring the present disclosure. References in this description to “an embodiment,” “one embodiment,” or the like, mean that a particular feature, structure, material, or characteristic being described is included in at least one embodiment of the present disclosure. Thus, the appearances of such phrases in this specification do not necessarily all refer to the same embodiment. On the other hand, such references are not necessarily mutually exclusive either. Furthermore, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments. Also, it is to be understood that the various embodiments shown in the figures are merely illustrative representations and are not necessarily drawn to scale.
In this patent document, the word “exemplary” is used to mean serving as an example, instance, or illustration. Any embodiment or design described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments or designs. Rather, use of the word exemplary is intended to present concepts in a concrete manner.
The 3D information of the surroundings is commonly stored as data in a format of point cloud—a set of data points representing actual locations of surrounding objects in a selected coordinate system.
The 3D information of the surroundings is then forwarded into a segmentation module to group the data points into various groups, each of the group corresponding to a surrounding object. The point cloud, as well as the results of segmentation (i.e., the groups), are fed into an object tracker 207. The object tracker 207 is operable to build models of target objects based on the point cloud of the surrounding objects, compute motion estimations for the target objects, and perform optimization to the models in order to minimize the effect of motion blur. Table 1 and
When the object tracker 207 initializes, it has zero target objects. Given some initial input data, it first identifies a target object that is deemed static with an initial motion estimation of Minit={0}. Upon receiving subsequent input St from the segmentation module 205, the object tracker 207 performs object identification, motion estimation, and optimization to obtain updated models for the target objects Pt,target at time t. Because the input date density from the LIDAR emitter-sensor pairs is relatively low, there could exist unidentified data points in St that cannot be mapped to any of the target objects. Such unidentified data points may be fed back to the segmentation module 205 for further segmentation at the next time t+1.
The object tracker 207 may include three separate components to complete the main steps shown in Table 1: an object identifier 211 that performs object identification, a motion estimator 213 that performs motion estimations, and an optimizer 215 that optimizes the models of the target objects. These components can be implemented in special-purpose computers or data processors that are specifically programmed, configured or constructed to perform the respective functionalities. Alternatively, an integrated component performing all these functionalities can also be implemented in a special-purpose computer or processor. Details regarding the functionalities of the object identifier 211, the motion estimator 213, and the optimizer 215 will be described in further details in connection with
The output of the object tracker 207, which includes models of target objects and the corresponding motion estimations, is then used by a control system 209 to facilitate decision making regarding the maneuver of the unmanned vehicle to avoid obstacles and to conduct adaptive cruising and/or lane switching.
P
t,target
′
=M
t-1,target
*P
t-1,target Eq. (1)
Based on the predicted locations of the target objects P′t,target and the actual locations of the surrounding objects Pt,surrounding, a similarity function co between the target objects and the surrounding objects can be evaluated, at 304, using a cost function F:
ωtarget,surrounding=F(Pt,target′,Pt,surrounding) Eq. (2)
The cost function F can be designed to accommodate specific cases. For example, F can simply be the center distance of the two point clouds P′t,target and Pt,surrounding, or the number of voxels commonly occupied by both P′t,target and Pt,surrounding. In some embodiments, the cost function F(P,Q) can be defined as:
F(P,Q)=Σp∈P∥p−q∥2, Eq. (3)
where p is a point in point cloud P and q is the closest point to point p in point cloud Q. The cost function F can also include color information for each point data supplied by the camera array 203, as shown in
Given the cost function F, a bipartite graph can be built, at 306, for all points contained in P′t,target and Pt,surrounding.
A complete bipartite graph can be built for all points in the target objects and all points in the surrounding objects. However, the computational complex of solving the complete bipartite graph is O(n{circumflex over ( )}3) where n is the number of objects. The performance can be substantially impacted when there is a large number of objects in the scene. To ensure the real time performance, subgraphs of the complete bipartite graph can be identified using the location information of the target object. This is based on an assumption that a target object is unlikely to undergo substantial movement between time t−1 and t. Its surface points are likely to located within a relative small range within the point cloud data set. Due to such locality of the data points, the complete bipartite graph can be divided into subgraphs. Each of the subgraph can be solved sequentially or concurrently using algorithms such as the KM algorithm.
After solving the bipartite graph (or subgraphs), the object tracker obtains, at 310, a mapping of the surrounding objects Pt,surrounding to the target objects Pt-1,target. In some cases, after solving the bipartite graph or subgraphs, not all target objects at time t-1 are map to objects in Pt,surrounding. This can happen when an object temporarily occluded by another object and becomes invisible to the LIDAR tracking system. For example, at time t, the object tracker cannot find a corresponding group within Pt,surrounding for the target object A. The object tracker considers the target object A still available and assigns a default motion estimation Mdefault to it. The object tracker further updates object A's model using Mdefault: Pt,A=Mdefault*Pt-1,A. Once the object becomes visible again, the system continues to track its locations. On the other hand, if the object tracker continuously fails to map any of the surrounding objects to the target object A for a predetermined amount of time, e.g., 1 second, the object tracker considers the target object A missing as if it has permanently moved outside of the sensing range of the LIDAR emitter-sensor pairs. The object tracker then deletes this particular target object from the models.
In some cases, not all surrounding objects Pt,surrounding in the input can be mapped to corresponding target objects. For example, the object tracker fails to map a group of points Bp in St, indicative of a surrounding object B, to any of the target objects Pt-1,target. To determine if the group of points Bp is a good representation of the object B, the object tracker evaluates the point density of Bp based on the amount of points in Bp and the distance from B to the LIDAR emitter-sensor pairs. For example, if the object B is close to the LIDAR emitter-sensor pairs, the object tracker requires more data points in Bp to be a sufficient representation of object B. On the other hand, if object B is far away from the LIDAR emitter-sensor pairs, even a small amount of data points in Bp may be sufficient to qualify as a good representation of object B. When the density is below a predetermined threshold, the object tracker 207 feeds the data points back to the segmentation module 205 for further segmentation at time t+1. On the other hand, if the group of data points has sufficient density and has presented in input data set for longer than a predetermined amount of time, e.g., 1 second, the object tracker 207 deems this group of points to be a new target object and initializes its states accordingly.
Motion Estimation
After object identification, the object tracker now obtains a mapping of Pt,surrounding to Pt-1,target.
Based on Pt-1,target and Pt,surrounding, the object tracker can compute a motion estimation Mt,target for time t.
T(Mt,Mt-1)=(Mt−μt-1)TΣt-1−1(Mt−μt-1) Eq. (4)
The constraint function T can describe uniform motion, acceleration, and rotation of the target objects. For example,
After the motion estimator 213 builds a model based on Mt-1, target, the motion estimation problem can essentially be described as solving an optimization problem defined as:
where λ is a parameter that balances the cost function F and the constraint function T. Because this optimization problem is highly constrained, the motion estimator 213 can discretize, at 604, the search of the Gaussian distribution model using the constraint function T as boundaries. The optimization problem is then transformed to a search problem for Mt. The motion estimator 213 then, at 606, searches for Mt within the search space defined by the discretized domain so that Mt minimizes:
F(Mt*Pt-1,Pt)+λT(Mt,Mt-1). Eq. (6)
In some embodiments, the motion estimator 213 can change the discretization step size adaptively based on density of the data points. For example, if object C is located closer to the LIDAR emitter-sensor pairs, the motion estimator 213 uses a dense discretization search scheme in order to achieve higher accuracy for the estimated results. If object D, on the other hand, is located further from the LIDAR emitter-sensor pairs, a larger discretization step size can be used for better search efficiency. Because evaluating Eq. (5) is mutually independent for each of the discretized step, in some embodiments, the search is performed concurrently on a multicore processor, such as a graphic processing unit (GPU), to increase search speed and facilitate real-time object tracking responses.
Lastly, after Mt,target is found in the discretized model, the motion estimator 213 updates, at 608, the point cloud models for the target objects based on the newly found motion estimation:
P
t,target
=M
t,target
*P
t-1,target Eq. (7)
Optimization
Because some of the target objects move at a very fast speed, a physical distortion, such as motion blur, may present in models for the target objects. The use of low-cost single-channel linear LIDAR emitter and sensor pairs may exacerbate this problem because, due to the low data density sensed by these LIDARs, it is desirable to have a longer accumulation time to accumulate sufficient data points for object classification and tracking. Longer accumulation time, however, means that there is a higher likelihood to encounter physical distortion in the input data set. An optimizer 215 can be implemented to reduce or remove the physical distortion in the models for the target objects and improve data accuracy for object tracking.
For example, for a particular point object E (that is, an object having only one point), n input data points, ρ0, ρ1, . . . , ρn-1 ∈Pt,surrounding are collected during the time Δt between t−1 and t. The data points are associated with timestamps defined as ti=t−(n−i)*Δt, where Δt is determined by the sensing frequency of the LIDAR emitter and sensor pairs. Subsequently, these data points are mapped to Pt-1,target. When the object tracker updates the model Pt,target for time t, the timestamps for ρ0, ρ1, . . . , ρn-1 are assigned to the corresponding points in the model Pt,target. These multiple input data points cause physical distortion of the point object D in Pt,target.
After the motion estimation Mt,target relative to the LIDAR system for time t is known, the absolution estimated motion for the target M_absolutet,target can be obtained using Mt,target and the speed of the LIDAR system. In some embodiments, the speed of the LIDAR system can be measured using an inertial measurement unit (IMU). Then, the optimizer 215, at 802, examines timestamps of each of the points in a target object Pt,target. For example, for the point object E, the accumulated point cloud data (with physical distortion) can be defined as:
U
i=0
n-1ρi Eq. (8)
The desired point cloud data (without physical distortion), however, can be defined as:
ρ=Ui=0n-1M_absolute′ti*ρi Eq. (9)
where M_absolute′ti is an adjusted motion estimation for each data point ρi at time ti. The optimizer 215 then, at 804, computes the adjusted motion estimation based on the timestamps of each point.
There are several ways to obtain the adjusted motion estimation M_absolute′ti. In some embodiments, M_absolute′ti can be computed by evaluating M_absolutet,target at different timestamps. For example, given M_absolutet,target, a velocity Vt,target of the target object can be computed. M_absolute′ti, therefore, can be calculated based on M_absolutet,target and (n−i)*Δt*Vt,target. Alternatively, a different optimization problem defined as follows can be solved to obtain M_absolute′ti:
where F′ can be defined in a variety of ways, such as the number of voxels ρ occupies. A similar discretized search method as described above can be applied to find the solution to M′.
Finally, after adjusting the motion estimation based on the timestamp, the optimizer 315 applies, at 806, the adjusted motion estimation to the corresponding data point to obtain a model with reduced physical distortion.
It is thus evident that, in one aspect of the disclosed technology, a light detection and ranging (LIDAR) based object tracking system. The system includes a plurality of light emitter and sensor pairs. Each pair of the plurality of light emitter and sensor pairs is operable to obtain data indicative of actual locations of surrounding objects. The data is grouped into a plurality of groups by a segmentation module, with each group corresponding to one of the surrounding objects. The system also includes an object tracker configured to (1) build a plurality of models of target objects based on the plurality of groups, (2) compute a motion estimation for each of the target objects, and (3) feed a subset of data back to the segmentation module for further classification based on a determination by the object tracker that the subset of data fails to map to a corresponding target object in the model.
In some embodiments, the object tracker includes an object identifier that (1) computes a predicted location for a target object among the target objects based on the motion estimation for the target object and (2) identifies, among the plurality of groups, a corresponding group that matches the target object. The object tracker also includes a motion estimator that updates the motion estimation for the target object by finding a set of translation and rotation values that, after applied to the target object, produces a smallest difference between the predicted location of the target object and the actual location of the corresponding group, wherein the motion estimator further updates the model for the target object using the motion estimation. The object tracker further includes an optimizer that modifies the model for the target object by adjusting the motion estimation to reduce or remove a physical distortion of the model for the target object.
In some embodiments, the object identifier identifies the corresponding group by evaluating a cost function, the cost function defined by a distance between the predicted location of the target object and the actual location of a group among the plurality of groups.
In some embodiments, the object tracking system further includes a camera array coupled to the plurality of light emitter and sensor pairs. The cost function is further defined by a color difference between the target object and the group, the color difference determined by color information captured by the camera array. The color information includes a one-component value or a three-component value in a predetermined color space.
In some embodiments, the object identifier identifies the corresponding group based on solving a complete bipartite graph of the cost function. In solving the complete bipartite graph, the object identifier can divide the complete bipartite graph to a plurality of subgraphs based on a location information of the target objects. The object identifier can solve the plurality of subgraphs based on a Kuhn-Munkres algorithm.
In some embodiments, the object identifier, upon determining that a target object fails to map to any of the actual locations of the surrounding objects for an amount of time no longer than a predetermined threshold, assigns the target object a uniform motion estimation. The object identifier may, upon determining that a target object fails to map to any of the actual locations of the surrounding objects for an amount of time longer than the predetermined threshold, remove the target object from the model.
In some embodiments, the object identifier, in response to a determination that the subset of data fails to map to any of the target objects, evaluates a density of the data in the subset, adds the subset as a new target object to the model when the density is above a predetermined threshold, and feeds the subset back to the segmentation module for further classification when the density is below the predetermined threshold.
In some embodiments, the motion estimator conducts a discretized search of a Gaussian motion model based on a set of predetermined, physics-based constraints of a given target object to compute the motion estimation. The system may further includes a multicore processor, wherein the motion estimator utilizes the multicore processor to conduct the discretized search of the Gaussian motion model in parallel. In some embodiments, the optimizer modifies the model for the target object by applying one or more adjusted motion estimations to the model.
In another aspect of the disclosed technology, a microcontroller system for controlling an unmanned movable object is disclosed. The system includes a processor configured to implement a method of tracking objects in real-time or near real-time. The method includes receiving data indicative of actual locations of surrounding objects. The actual locations are classified into a plurality of groups by a segmentation module, and each group of the plurality of groups corresponds to one of the surrounding objects. The method also includes obtaining a plurality of models of target objects based on the plurality of groups; estimating a motion matrix for each of the target objects; updating the model using the motion matrix for each of the target objects; and optimizing the model by modifying the model for each of the target objects to remove or reduce a physical distortion of the model for the target object.
In some embodiments, the obtaining of the plurality of models of the target objects includes computing a predicted location for each of the target objects; and identifying, based on the predicted point location, a corresponding group among the plurality of groups that maps to a target object among the target objects. The identifying of the corresponding group can include evaluating a cost function that is defined by a distance between the predicted location of the target object and the actual location of a group among the plurality of groups.
In some embodiments, the system further includes a camera array coupled to the plurality of light emitter and sensor pairs. The cost function is further defined by a color difference between the target object and the group, the color difference determined by color information captured by a camera array. The color information may include a one-component value or a three-component value in a pre-determined color space.
In some embodiments, the identifying comprises solving a complete bipartite graph of the cost function. In solving the complete bipartite graph, the processor divides the complete bipartite graph to a plurality of subgraphs based on a location information of the target objects. The processor can solve the plurality of subgraphs using a Kuhn-Munkres algorithm.
In some embodiments, the identifying comprises assigning a target object a uniform motion matrix in response to a determination that that the target object fails to map to any of the actual locations of the surrounding objects for an amount of time shorter than a predetermined threshold. The identifying may include removing a target object from the model in response to a determination that the target object fails to map to any of the actual locations of the surrounding objects for an amount of time longer than the predetermined threshold. The identifying may also include, in response to a determination that a subset of the data fails to map to any of the target objects, evaluating a density of data in the subset, adding the subset as a new target object if the density is above a predetermined threshold, and feeding the subset back to the segmentation module for further classification based on a determination that the density is below the predetermined threshold.
In some embodiments, the estimating includes conducting a discretized search of a Gaussian motion model based on a set of prior constraints to estimate the motion matrix, wherein a step size of the discretized search is determined adaptively based on a distance of each of the target objects to the microcontroller system. The conducting can include subdividing the discretized search of the Gaussian motion model into sub-searches and conducting the sub-searches in parallel on a multicore processor.
In some embodiments, the optimizing includes evaluating a velocity of each of the target objects, and determining, based on the evaluation, whether to apply one or more adjusted motion matrices to the target object to remove or reduce the physical distortion of the model.
In yet another aspect of the disclosed technology, an unmanned device is disclosed. The unmanned device comprises a light detection and ranging (LIDAR) based object tracking system as described above, a controller operable to generate control signals to direct motion of the vehicle in response to output from the real-time object tracking system, and an engine operable to maneuver the vehicle in response to control signals from the controller.
Some of the embodiments described herein are described in the general context of methods or processes, which may be implemented in one embodiment by a computer program product, embodied in a computer-readable medium, including computer-executable instructions, such as program code, executed by computers in networked environments. A computer-readable medium may include removable and non-removable storage devices including, but not limited to, Read Only Memory (ROM), Random Access Memory (RAM), compact discs (CDs), digital versatile discs (DVD), etc. Therefore, the computer-readable media can include a non-transitory storage media. Generally, program modules may include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer- or processor-executable instructions, associated data structures, and program modules represent examples of program code for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps or processes.
Some of the disclosed embodiments can be implemented as devices or modules using hardware circuits, software, or combinations thereof. For example, a hardware circuit implementation can include discrete analog and/or digital components that are, for example, integrated as part of a printed circuit board. Alternatively, or additionally, the disclosed components or modules can be implemented as an Application Specific Integrated Circuit (ASIC) and/or as a Field Programmable Gate Array (FPGA) device. Some implementations may additionally or alternatively include a digital signal processor (DSP) that is a specialized microprocessor with an architecture optimized for the operational needs of digital signal processing associated with the disclosed functionalities of this application. Similarly, the various components or sub-components within each module may be implemented in software, hardware or firmware. The connectivity between the modules and/or components within the modules may be provided using any one of the connectivity methods and media that is known in the art, including, but not limited to, communications over the Internet, wired, or wireless networks using the appropriate protocols.
While this patent document contains many specifics, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this patent document in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.
Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. Moreover, the separation of various system components in the embodiments described in this patent document should not be understood as requiring such separation in all embodiments.
Only a few implementations and examples are described and other implementations, enhancements and variations can be made based on what is described and illustrated in this patent document.
| Number | Date | Country | Kind |
|---|---|---|---|
| PCT/CN2017/082601 | Apr 2017 | CN | national |
This application is a continuation of International Application No. PCT/CN2017/110534, filed Nov. 10, 2017, which claims priority to International Application No. PCT/CN2017/082601, filed Apr. 28, 2017, the entire contents of both of which are incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/CN2017/110534 | Nov 2017 | US |
| Child | 16664331 | US |
