Detection transformer (DETR) back propagation using global loss function represented as sum of allocation independent terms and allocation dependent terms defined by allocation cost matrix

Abstract

In an embodiment, a method for training a detection transducer (DETR) includes initializing, with at least one processor, a parameter of the DETR; propagating, with the at least one processor, an input image through the DETR; determining, with the at least one processor, an error value by comparing an output of the DETR with a known expected output; and iteratively updating, with the at least one processor, a parameter in the DETR based on the error value by minimizing a global loss, where minimizing the global loss includes: generating an allocation cost matrix; solving an allocation optimization problem based on the allocation cost matrix to obtain an optimal allocation cost; a global loss function is then minimized, which is represented as a sum of the optimal allocation cost and the allocation-independent loss.

Description

[0001] Cross - reference to related applications

[0002] This application claims the benefit of U.S. Provisional Patent Application Serial No. 63 / 416,481, filed Oct. 14, 2022, and U.S. Provisional Patent Application Serial No. 63 / 424,831, filed Nov. 11, 2022, under 35 U.S.C. § 119(e), the entire contents of both of which are hereby incorporated by reference. Background Art

[0003] The Detection Transformer (DETR), along with its subsequent variants such as Deformable DETR, etc., has become a building block for many transformer - based object detection and tracking methods used in autonomous vehicles and other applications. DETR uses a transformer encoder - decoder architecture and a set - based global loss that enforces unique predictions using bipartite matching. In traditional DETR methods, the assignment cost and the global loss are inconsistent, i.e., it is possible to reduce the assignment cost without guaranteeing a reduction in the global loss. Additionally, when using a combinatorial solver such as the Hungarian algorithm, the issue of gradients is ignored. Brief Description of the Drawings

[0004] Figure 1 is an example environment of a vehicle that can implement one or more components of an autonomous system;

[0005] Figure 2 is a diagram of one or more systems of a vehicle including an autonomous system;

[0006] Figure 3 is Figure 1 and Figure 2 is a diagram of one or more devices and / or components of one or more systems;

[0007] Figure 4A is a diagram of certain components of an autonomous system;

[0008] Figure 4B is a diagram of an implementation of a neural network;

[0009] Figure 4C and Figure 4D is a diagram illustrating an example operation of a CNN;

[0010] Figure 5 is a block diagram of a DETR transformer architecture; and

[0011] Figure 6 is a flowchart of DETR transformer processing. Detailed Description

[0012] In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. It will be apparent, however, that the embodiments described herein may be practiced without these specific details. In some instances, well-known structures and devices are illustrated in block diagram form in order to avoid unnecessarily obscuring aspects of the present disclosure.

[0013] In the drawings, for ease of description, the specific arrangements or orderings of illustrative elements (such as those representing systems, devices, modules, instruction blocks, and / or data elements, etc.) are illustrated. However, those skilled in the art will understand that, unless explicitly described, the specific orderings or arrangements of the illustrative elements in the drawings are not intended to imply a requirement for a particular processing order or sequence, or a separation of processes. Additionally, unless explicitly described, the inclusion of illustrative elements in the drawings is not intended to imply that such elements are required in all embodiments, nor that the features represented by such elements cannot be included in some embodiments or combined with other elements in some embodiments.

[0014] Furthermore, in the drawings, connecting elements (such as solid lines, dashed lines, or arrows, etc.) are used to illustrate connections, relationships, or associations between or among two or more other illustrative elements. The absence of any such connecting element is not intended to imply that connections, relationships, or associations cannot exist. In other words, some connections, relationships, or associations between elements are not illustrated in the drawings so as not to obscure the present disclosure. Additionally, for ease of illustration, a single connecting element may be used to represent multiple connections, relationships, or associations between elements. For example, if a connecting element represents the communication of a signal, data, or instruction (e.g., "software instruction"), those skilled in the art should understand that such an element may represent one or more signal paths (e.g., a bus) that may be required to affect the communication.

[0015] Although terms such as "first," "second," and / or "third" are used to describe various elements, these elements should not be limited by these terms. The terms "first," "second," and / or "third" are only used to distinguish one element from another. For example, without departing from the scope of the described embodiments, a first contact may be referred to as a second contact, and similarly, a second contact may be referred to as a first contact. Both the first contact and the second contact are contacts, but they are not the same contact.

[0016] The terms used in the description of the various embodiments described herein are included for the purpose of describing particular embodiments only and are not intended to be limiting. As used in the description of the various embodiments and the appended claims, the singular forms "a", "an", and "the" are also intended to include the plural forms and may be used interchangeably with "one or more than one" or "at least one", unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items. It will also be understood that when the terms "comprises", "comprising", "includes", and / or "including" are used in this specification, specify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0017] As used herein, the terms "communicate" and "communicating" refer to at least one of receiving, receiving, transmitting, conveying, and / or providing information (or information represented by, for example, data, signals, messages, instructions, and / or commands, etc.). For a unit (e.g., a device, a system, a component of a device or system, and / or a combination thereof, etc.) that is to communicate with another unit, this means that the one unit is capable of directly or indirectly receiving information from the other unit and / or sending (e.g., transmitting) information to the other unit. This may refer to a direct or indirect connection that is inherently wired and / or wireless. Additionally, two units may communicate with each other even if the information transmitted between the first unit and the second unit is modified, processed, relayed, and / or routed. For example, even if the first unit receives information passively and does not actively transmit information to the second unit, the first unit may communicate with the second unit. As another example, if at least one intermediate unit (e.g., a third unit located between the first unit and the second unit) processes the information received from the first unit and transmits the processed information to the second unit, the first unit may communicate with the second unit. In some embodiments, a message may refer to a network packet (e.g., a data packet, etc.) that includes data.

[0018] As used herein, depending on the context, the term "if" is optionally interpreted to mean "when", "upon", "in response to determining that", and / or "in response to detecting", etc. Similarly, depending on the context, the phrase "if it has been determined" or "if [stated condition or event] is detected" is optionally interpreted to mean "when determining...", "in response to determining that", or "when [stated condition or event] is detected" and / or "in response to detecting [stated condition or event]", etc. Further, as used herein, the terms "have", "having", or "possess", etc. are intended to be open-ended terms. Additionally, unless expressly stated otherwise, the phrase "based on" is intended to mean "at least partially based on".

[0019] Reference will now be made in detail to the embodiments, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the various described embodiments. However, it will be apparent to one of ordinary skill in the art that the various described embodiments may be practiced without these specific details. In other instances, well-known methods, processes, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the embodiments.

[0020] General Overview

[0021] Autonomous vehicles rely on perception of their surrounding environment to ensure safe and robust driving performance. This perception includes the ability to simultaneously detect and track multiple objects. One technology for perception is DETR, which uses transformers and bipartite matching loss for direct set prediction. DETR was first described in Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. volume 12346 LNCS, pages 213-229. Springer Science and Business Media Deutschland GmbH, 2020 (hereinafter referred to as the "pioneering paper").

[0022] In this pioneering paper, the network produces a set of predictions that is larger in number than the number of ground truth boxes. The assignment problem is defined and solved by a Hungarian algorithm solver, and the globally optimal matching boxes are used to define the global loss to be minimized by backpropagation. The global loss takes into account three aspects of the predictions: (1) the class of the matched box should be the class of its assigned ground truth; (2) the position and size of these boxes should be those of their assigned ground truth; and (3) the class of the unmatched boxes should be the background.

[0023] Intuitively, the criterion according to which the matching is performed (i.e., the total assignment cost) should be consistent with the global loss, such that reducing the total assignment cost must reduce the global loss. However, this is not the case because the matching cost is defined in a different way from the global loss. The term considering the class uses the raw probability instead of cross entropy. The heuristic reason given is the relative proportion of the loss from geometry. One might think that the scaling problem is the same as in the loss and that an additional scaling hyperparameter can be used to explicitly address it. The above term is also restricted to the matched boxes only. The reason given is the wrong reason because each prediction has a different probability of being the background and thus the missing sum is not matching-independent.

[0024] In a broader context, the Hungarian algorithm solver can be regarded simply as another module that performs some operations (in this case, some discrete optimization). One might think that it does not matter how the cost matrix is defined as long as the gradients can be properly taken into account. Unfortunately, this is not achieved in most DETR methods. In the code released together with the pioneering paper, the problem of gradients is ignored by wrapping the matcher code with torch.no_grad(), i.e., gradient tracking is turned off when the Hungarian algorithm solver is involved.

[0025] The disclosed embodiments provide an alternative and simpler approach. First, the global loss is expressed as the sum of two terms: the first term is defined by the probability of all predictions being the background regardless of the matching. If the cost matrix is properly defined, the second term is regarded as the optimal cost of the matching.

[0026] Thus, to backpropagate the global loss, it is only necessary to determine what the gradient of the optimal cost is with respect to the parameters defining the assignment problem. Fortunately, as described in Xi Gao, Han Zhang, Aliakbar Panahi, and Tom Arodz. Combinatorial losses through generalized gradients of integer linear programs. arXiv preprint arXiv:1910.08211, 2019, this can be determined using integer linear programming (ILP).

[0027] In some embodiments, a method for training a detection transformer, i.e., DETR, includes: initializing parameters of the DETR using at least one processor; propagating an input image through the DETR using the at least one processor; determining an error value by comparing an output of the DETR with a known expected output using the at least one processor; and iteratively updating the parameters in the DETR based on the error value by minimizing a global loss, where minimizing the global loss includes: generating an assignment cost matrix; solving an assignment optimization problem based on the assignment cost matrix to obtain an optimal assignment cost; and then minimizing a global loss function represented as a sum of the optimal assignment cost and an assignment-independent loss.

[0028] In some embodiments, the assignment cost matrix is a rectangular matrix.

[0029] In some embodiments, the assignment cost matrix is defined by ground truth annotations and network predictions as a function of network weight operators.

[0030] In some embodiments, the method further includes: filling the assignment cost matrix with random values that are nearly zero to form a square matrix.

[0031] In some embodiments, the method further includes: subtracting cross-entropy losses corresponding to the background from each row of the assignment cost matrix.

[0032] In some embodiments, the method further includes: using a Hungarian algorithm solver to solve the assignment problem.

[0033] In some embodiments, the assignment-independent term is an unmatched prediction loss.

[0034] In some embodiments, an object detector includes: a convolutional layer; a transformer encoder; a transformer decoder; and at least one prediction head, wherein the parameters of the detection transformer are determined by: propagating an input image through the object detector; determining an error value by comparing the output of the at least one prediction head with a known expected output; and iteratively updating the parameters in the object detector based on the error value by minimizing a global loss, wherein minimizing the global loss includes: generating an assignment cost matrix; using the assignment cost matrix to solve an assignment optimization problem to obtain an optimal assignment cost; and then minimizing a global loss function, which is expressed as the sum of the optimal assignment cost and an assignment-independent loss.

[0035] In some embodiments, a method includes: obtaining image data from at least one sensor of a vehicle; using at least one processor to detect at least one object captured in the image data, wherein the detection includes: inputting the image data into a detection transformer, i.e., DETR, having parameters determined during training by minimizing a global loss, wherein minimizing the global loss includes: generating an assignment cost matrix; solving an assignment optimization problem based on the assignment cost matrix to obtain an optimal assignment cost; and then minimizing a global loss function, which is expressed as the sum of the optimal assignment cost and an assignment-independent loss; and using the at least one processor to generate at least one control signal for controlling the vehicle, at least in part based on the detected at least one object.

[0036] By means of the embodiments described herein, the disclosed systems and methods improve the speed and accuracy of training a DETR detector relative to traditional training forms of DETR.

[0037] Now refer to Figure 1, an exemplary environment 100 is illustrated, in which vehicles including autonomous systems and vehicles not including autonomous systems operate. As illustrated, environment 100 includes vehicles 102a - 102n, objects 104a - 104n, routes 106a - 106n, area 108, vehicle - to - infrastructure (V2I) devices 110, network 112, remote autonomous vehicle (AV) system 114, queue management system 116, and V2I system 118. Vehicles 102a - 102n, vehicle - to - infrastructure (V2I) devices 110, network 112, autonomous vehicle (AV) system 114, queue management system 116, and V2I system 118 are interconnected via a wired connection, a wireless connection, or a combination of wired and wireless connections (e.g., establishing a connection for communication, etc.). In some embodiments, objects 104a - 104n are interconnected with at least one of vehicles 102a - 102n, vehicle - to - infrastructure (V2I) devices 110, network 112, autonomous vehicle (AV) system 114, queue management system 116, and V2I system 118 via a wired connection, a wireless connection, or a combination of wired and wireless connections.

[0038] Vehicles 102a - 102n (individually referred to as vehicle 102 and collectively as vehicles 102) include at least one device configured to transport goods and / or people. In some embodiments, vehicle 102 is configured to communicate with V2I device 110, remote AV system 114, queue management system 116, and / or V2I system 118 via network 112. In some embodiments, vehicles 102 include cars, buses, trucks, and / or trains, etc. In some embodiments, vehicle 102 is the same as or similar to vehicle 200 described herein (see Figure 2 ). In some embodiments, vehicles 200 in the set of vehicles 200 are associated with an autonomous queue manager. In some embodiments, as described herein, vehicle 102 travels along corresponding routes 106a - 106n (individually referred to as route 106 and collectively as routes 106). In some embodiments, one or more than one vehicle 102 includes an autonomous system (e.g., an autonomous system the same as or similar to autonomous system 202).

[0039] Objects 104a - 104n (individually referred to as object 104 and collectively as objects 104) include, for example, at least one vehicle, at least one pedestrian, at least one cyclist, and / or at least one structure (e.g., building, sign, fire hydrant, etc.). Each object 104 (e.g., located at a fixed location and over a period of time) is stationary or (e.g., having a speed and associated with at least one trajectory) moving. In some embodiments, object 104 is associated with a corresponding location in region 108.

[0040] Routes 106a - 106n (individually referred to as route 106 and collectively as routes 106) are each associated with (e.g., define) a sequence of actions (also referred to as a trajectory) that a connected AV can navigate along. Each route 106 begins at an initial state (e.g., a state corresponding to a first spatio - temporal location and / or speed, etc.) and ends at a final target state (e.g., a state corresponding to a second spatio - temporal location different from the first spatio - temporal location) or a target zone (e.g., a subspace of acceptable states (e.g., termination states)). In some embodiments, the first state includes a location where one or more individuals will board the AV, and the second state or zone includes one or more locations where one or more individuals boarding the AV will disembark. In some embodiments, route 106 includes multiple acceptable sequences of states (e.g., multiple sequences of spatio - temporal locations) that are associated with (e.g., define) multiple trajectories. In an example, route 106 includes only high - level actions or imprecise state locations, such as a series of connected roads indicating a direction change at a roadway intersection. Additionally or alternatively, route 106 can include more precise actions or states, such as, for example, a specific target lane or precise location within a lane region and a target rate at those locations. In an example, route 106 includes multiple precise state sequences along at least one high - level action with a finite look - ahead horizon to reach an intermediate target, where the combination of successive iterations of the finite - horizon state sequences cumulatively corresponds to multiple trajectories that together form a high - level route terminating at the final target state or zone.

[0041] Region 108 includes a physical region (e.g., a geographic region) in which vehicle 102 can navigate. In an example, region 108 includes at least one state (e.g., a country, a province, an individual state among multiple states included in a country, etc.), at least a portion of a state, at least one city, at least a portion of a city, etc. In some embodiments, region 108 includes at least one named arterial road (referred to herein as a "road"), such as a highway, an interstate highway, a parkway, a city street, etc. Additionally or alternatively, in some examples, region 108 includes at least one unnamed road, such as a lane, a section of a parking lot, a section of a vacant and / or undeveloped area, a dirt road, etc. In some embodiments, a road includes at least one lane (e.g., a portion of the road through which vehicle 102 can pass). In an example, a road includes at least one lane associated with (e.g., identified based on) at least one lane marking line.

[0042] A vehicle-to-infrastructure (V2I) device 110 (sometimes referred to as a vehicle-to-infrastructure or vehicle-to-everything (V2X) device) includes at least one device configured to communicate with vehicle 102 and / or V2I system 118. In some embodiments, V2I device 110 is configured to communicate with vehicle 102, remote AV system 114, queue management system 116, and / or V2I system 118 via network 112. In some embodiments, V2I device 110 includes a radio frequency identification (RFID) device, a sign, a camera (e.g., a two-dimensional (2D) and / or three-dimensional (3D) camera), a lane marking, a street light, a parking meter, etc. In some embodiments, V2I device 110 is configured to communicate directly with vehicle 102. Additionally or alternatively, in some embodiments, V2I device 110 is configured to communicate with vehicle 102, remote AV system 114, and / or queue management system 116 via V2I system 118. In some embodiments, V2I device 110 is configured to communicate with V2I system 118 via network 112.

[0043] Network 112 includes one or more wired and / or wireless networks. In an example, network 112 includes a cellular network (e.g., a Long-Term Evolution (LTE) network, a third-generation (3G) network, a fourth-generation (4G) network, a fifth-generation (5G) network, a Code Division Multiple Access (CDMA) network, etc.), a Public Land Mobile Network (PLMN), a Local Area Network (LAN), a Wide Area Network (WAN), a Metropolitan Area Network (MAN), a telephone network (e.g., a Public Switched Telephone Network (PSTN)), a private network, an ad hoc network, an intranet, the Internet, a fiber-based network, a cloud computing network, etc., and / or a combination of some or all of these networks.

[0044] The remote AV system 114 includes at least one device configured to communicate with the vehicle 102, the V2I device 110, the network 112, the queue management system 116, and / or the V2I system 118 via the network 112. In an example, the remote AV system 114 includes a server, a server group, and / or other similar devices. In some embodiments, the remote AV system 114 is co-located with the queue management system 116. In some embodiments, the remote AV system 114 participates in the installation of some or all of the components of the vehicle (including autonomous systems, autonomous vehicle computing, and / or software implemented by autonomous vehicle computing, etc.). In some embodiments, the remote AV system 114 maintains (e.g., updates and / or replaces) these components and / or software during the life of the vehicle.

[0045] The queue management system 116 includes at least one device configured to communicate with the vehicle 102, the V2I device 110, the remote AV system 114, and / or the V2I system 118. In an example, the queue management system 116 includes a server, a server group, and / or other similar devices. In some embodiments, the queue management system 116 is associated with a ridesharing company (e.g., an organization for controlling the operation of multiple vehicles (e.g., vehicles including autonomous systems and / or vehicles not including autonomous systems), etc.).

[0046] In some embodiments, the V2I system 118 includes at least one device configured to communicate with the vehicle 102, the V2I device 110, the remote AV system 114, and / or the queue management system 116 via the network 112. In some examples, the V2I system 118 is configured to communicate with the V2I device 110 via a connection different from the network 112. In some embodiments, the V2I system 118 includes a server, a server group, and / or other similar devices. In some embodiments, the V2I system 118 is associated with a municipal authority or a private institution (e.g., a private institution for maintaining the V2I device 110, etc.).

[0047] Provided Figure 1 The number and arrangement of the illustrated elements are provided as examples. Compared with Figure 1 the illustrated elements, there may be additional elements, fewer elements, different elements, and / or elements with a different arrangement. Additionally or alternatively, at least one element of the environment 100 may perform one or more functions described as being performed by Figure 1 at least one different element. Additionally or alternatively, at least one set of elements of the environment 100 may perform one or more functions described as being performed by at least one different set of elements of the environment 100.

[0048] Now refer to Figure 2 , vehicle 200 (which may be the same as or similar to Figure 1 's vehicle 102) includes an autonomous system 202, a powertrain control system 204, a steering control system 206, and a braking system 208, or is associated with the autonomous system 202, the powertrain control system 204, the steering control system 206, and the braking system 208. In some embodiments, vehicle 200 is the same as or similar to vehicle 102 (see Figure 1 ). In some embodiments, the autonomous system 202 is configured to endow vehicle 200 with autonomous driving capabilities (e.g., implement at least one of the following driving functions, features, and / or devices that are automatic or based on maneuvers, and the at least one driving function, feature, and / or device that is automatic or based on maneuvers enables vehicle 200 to operate partially or completely without human intervention, including but not limited to fully autonomous vehicles (e.g., vehicles that abandon reliance on human intervention, such as level 5 ADS-operated vehicles), highly autonomous vehicles (e.g., vehicles that abandon reliance on human intervention in certain situations, such as level 4 ADS-operated vehicles), and / or conditionally autonomous vehicles (e.g., vehicles that abandon reliance on human intervention in limited situations, such as level 3 ADS-operated vehicles), etc.). In one embodiment, the autonomous system 202 includes the operational or tactical functions required to enable vehicle 200 to operate in road traffic and continuously perform part or all of the dynamic driving task (DDT). In another embodiment, the autonomous system 202 includes an advanced driver assistance system (ADAS) that includes driver support features. The autonomous system 202 supports various levels of driving automation ranging from no driving automation (e.g., level 0) to full driving automation (e.g., level 5). For a detailed description of fully autonomous vehicles and highly autonomous vehicles, reference can be made to SAE International Standard J3016: Taxonomy and Definitions for Terms Related to On-Road Motor Vehicle Automated Driving Systems, the entire content of which is incorporated by reference. In some embodiments, vehicle 200 is associated with an autonomous queue manager and / or a ridesharing company.

[0049] The autonomous system 202 includes a sensor suite that includes one or more devices such as a camera 202a, a LiDAR sensor 202b, a Radar sensor 202c, and a microphone 202d. In some embodiments, the autonomous system 202 may include more or fewer devices and / or different devices (e.g., ultrasonic sensors, inertial sensors, GPS receivers (discussed below), and / or odometer sensors for generating data associated with an indication of the distance traveled by the vehicle 200, etc.). In some embodiments, the autonomous system 202 uses one or more devices included in the autonomous system 202 to generate data associated with the environment 100 described herein. The data generated by one or more devices of the autonomous system 202 may be used by one or more systems described herein to observe the environment (e.g., environment 100) in which the vehicle 200 is located. In some embodiments, the autonomous system 202 includes a communication device 202e, an autonomous vehicle computing 202f, a drive-by-wire (DBW) system 202h, and a safety controller 202g.

[0050] The camera 202a includes at least one device configured to communicate with the communication device 202e, the autonomous vehicle computing 202f, and / or the safety controller 202g via a bus (e.g., a bus 302 that is the same as or similar to Figure 3 the bus). The camera 202a includes at least one camera (e.g., a digital camera using an optical sensor such as a charge-coupled device (CCD), a thermal camera, an infrared (IR) camera, and / or an event camera, etc.) for capturing images including physical objects (e.g., cars, buses, curbs, and / or people, etc.). In some embodiments, the camera 202a generates camera data as output. In some examples, the camera 202a generates camera data including image data associated with the image. In this example, the image data may specify at least one parameter corresponding to the image (e.g., image characteristics such as exposure, brightness, etc., and / or an image timestamp, etc.). In such an example, the image may be in a format (e.g., RAW, JPEG, and / or PNG, etc.). In some embodiments, the camera 202a includes a plurality of independent cameras configured (e.g., positioned) on the vehicle for capturing images for the purpose of stereovision (stereo vision). In some examples, the camera 202a generates image data and transmits the image data to the autonomous vehicle computing 202f and / or a queue management system (e.g., the same as Figure 1A plurality of cameras of the same or similar queue management system as the queue management system 116. In such an example, the autonomous vehicle computing 202f determines the depth to one or more objects in the fields of view of at least two of the plurality of cameras based on image data from at least two cameras. In some embodiments, the camera 202a is configured to capture images of objects within a distance relative to the camera 202a (e.g., up to 100 meters and / or up to 1 kilometer, etc.). Thus, the camera 202a includes features such as sensors and lenses optimized for sensing objects at one or more distances relative to the camera 202a.

[0051] In an embodiment, the camera 202a includes at least one camera configured to capture one or more images associated with one or more traffic lights, street signs, and / or other physical objects providing visual navigation information. In some embodiments, the camera 202a generates traffic light data associated with one or more images. In some examples, the camera 202a generates TLD (Traffic Light Detection) data associated with one or more images including a format (e.g., RAW, JPEG, and / or PNG, etc.). In some embodiments, the camera 202a that generates TLD data is different from other systems incorporating cameras described herein in that: the camera 202a may include one or more cameras having a wide field of view (e.g., a wide-angle lens, a fish-eye lens, and / or a lens having a viewing angle of about 120 degrees or greater, etc.) to generate images related to as many physical objects as possible.

[0052] The Light Detection and Ranging (LiDAR) sensor 202b includes being configured to communicate with the communication device 202e, the autonomous vehicle computing 202f, and / or the safety controller 202g via a bus (e.g., with Figure 3at least one device that communicates via a bus (e.g., a bus identical or similar to bus 302). The LiDAR sensor 202b includes a system configured to emit light from a light emitter (e.g., a laser emitter). The light emitted by the LiDAR sensor 202b includes light outside the visible spectrum (e.g., infrared light, etc.). In some embodiments, during operation, the light emitted by the LiDAR sensor 202b encounters a physical object (e.g., a vehicle) and is reflected back to the LiDAR sensor 202b. In some embodiments, the light emitted by the LiDAR sensor 202b does not penetrate the physical object that the light encounters. The LiDAR sensor 202b also includes at least one light detector that detects the light after the light emitted from the light emitter encounters a physical object. In some embodiments, at least one data processing system associated with the LiDAR sensor 202b generates an image (e.g., a point cloud and / or a combined point cloud, etc.) representing the objects included in the field of view of the LiDAR sensor 202b. In some examples, at least one data processing system associated with the LiDAR sensor 202b generates an image representing the boundary of a physical object and / or the surface of a physical object (e.g., the topology of the surface), etc. In such examples, the image is used to determine the boundary of the physical object in the field of view of the LiDAR sensor 202b.

[0053] A Radio Detection and Ranging (Radar) sensor 202c includes at least one device configured to communicate with a communication device 202e, an autonomous vehicle computer 202f, and / or a safety controller 202g via a bus (e.g., a bus identical or similar to Figure 3 bus 302). The Radar sensor 202c includes a system configured to emit (pulsed or continuous) radio waves. The radio waves emitted by the Radar sensor 202c include radio waves within a predetermined spectrum. In some embodiments, during operation, the radio waves emitted by the Radar sensor 202c encounter a physical object and are reflected back to the Radar sensor 202c. In some embodiments, the radio waves emitted by the Radar sensor 202c are not reflected by some objects. In some embodiments, at least one data processing system associated with the Radar sensor 202c generates a signal representing the objects included in the field of view of the Radar sensor 202c. For example, at least one data processing system associated with the Radar sensor 202c generates an image representing the boundary of a physical object and / or the surface of a physical object (e.g., the topology of the surface), etc. In some examples, the image is used to determine the boundary of the physical object in the field of view of the Radar sensor 202c.

[0054] The microphone 202d includes at least one device configured to communicate with the communication device 202e, the autonomous vehicle computing 202f, and / or the safety controller 202g via a bus (e.g., a bus the same as or similar to the Figure 3 bus 302). The microphone 202d includes one or more microphones (e.g., an array microphone and / or an external microphone, etc.) that capture an audio signal and generate data associated with (e.g., representing) the audio signal. In some examples, the microphone 202d includes a transducer device and / or a similar device. In some embodiments, one or more of the systems described herein can receive the data generated by the microphone 202d and determine the position (e.g., distance, etc.) of an object relative to the vehicle 200 based on the audio signal associated with the data.

[0055] The communication device 202e includes at least one device configured to communicate with the camera 202a, the LiDAR sensor 202b, the Radar sensor 202c, the microphone 202d, the autonomous vehicle computing 202f, the safety controller 202g, and / or the DBW (drive-by-wire) system 202h. For example, the communication device 202e can include a device the same as or similar to the Figure 3 communication interface 314. In some embodiments, the communication device 202e includes a vehicle-to-vehicle (V2V) communication device (e.g., a device for enabling wireless communication of data between vehicles).

[0056] The autonomous vehicle computing 202f includes at least one device configured to communicate with the camera 202a, the LiDAR sensor 202b, the Radar sensor 202c, the microphone 202d, the communication device 202e, the safety controller 202g, and / or the DBW system 202h. In some examples, the autonomous vehicle computing 202f includes devices such as a client device, a mobile device (e.g., a cellular phone and / or a tablet, etc.), and / or a server (e.g., a computing device including one or more central processing units and / or graphics processing units, etc.). In some embodiments, the autonomous vehicle computing 202f is configured to implement the autonomous vehicle software 400 described herein. In an embodiment, the autonomous vehicle computing 202f is the same as or similar to the DETR architecture 500 described herein. Additionally or alternatively, in some embodiments, the autonomous vehicle computing 202f is configured to communicate with an autonomous vehicle system (e.g., an autonomous vehicle system the same as or similar to the Figure 1 remote AV system 114), a queue management system (e.g., a queue management system the same as or similar to the Figure 1 queue management system 116), a V2I device (e.g., a device the same as or similar to the Figure 1the same or similar V2I device as the V2I device 110) and / or a V2I system (e.g., a V2I system the same or similar to the V2I system 118) Figure 1 communicate with the same or similar V2I system as the V2I system 118).

[0057] The safety controller 202g includes at least one device configured to communicate with the camera 202a, the LiDAR sensor 202b, the Radar sensor 202c, the microphone 202d, the communication device 202e, the autonomous vehicle computing 202f, and / or the DBW system 202h. In some examples, the safety controller 202g includes one or more controllers (such as an electrical controller and / or an electromechanical controller, etc.) configured to generate and / or transmit control signals to operate one or more devices of the vehicle 200 (such as the powertrain control system 204, the steering control system 206, and / or the braking system 208, etc.). In some embodiments, the safety controller 202g is configured to generate control signals that take precedence over (e.g., override) the control signals generated and / or transmitted by the autonomous vehicle computing 202f.

[0058] The DBW system 202h includes at least one device configured to communicate with the communication device 202e and / or the autonomous vehicle computing 202f. In some examples, the DBW system 202h includes one or more controllers (such as an electrical controller and / or an electromechanical controller, etc.) configured to generate and / or transmit control signals to operate one or more devices of the vehicle 200 (such as the powertrain control system 204, the steering control system 206, and / or the braking system 208, etc.). Additionally or alternatively, one or more controllers of the DBW system 202h are configured to generate and / or transmit control signals to operate at least one different device of the vehicle 200 (such as turn signals, headlights, door locks, and / or windshield wipers, etc.).

[0059] The powertrain control system 204 includes at least one device configured to communicate with the DBW system 202h. In some examples, the powertrain control system 204 includes at least one controller and / or actuator, etc. In some embodiments, the powertrain control system 204 receives control signals from the DBW system 202h, and the powertrain control system 204 causes the vehicle 200 to perform longitudinal vehicle movements (such as starting to move forward, stopping moving forward, starting to move backward, stopping moving backward, accelerating in a certain direction, decelerating in a certain direction, etc.) or perform lateral vehicle movements (such as making a left turn and / or making a right turn, etc.). In an example, the powertrain control system 204 increases, maintains the same, or decreases the energy (such as fuel and / or electricity, etc.) provided to the motor of the vehicle, thereby causing at least one wheel of the vehicle 200 to rotate or not rotate.

[0060] The steering control system 206 includes at least one device configured to rotate one or more wheels of the vehicle 200. In some examples, the steering control system 206 includes at least one controller and / or actuator, etc. In some embodiments, the steering control system 206 rotates two front wheels and / or two rear wheels of the vehicle 200 left or right to turn the vehicle 200 left or right. In other words, the steering control system 206 causes the activities required to regulate the y-axis component of the vehicle's movement.

[0061] The braking system 208 includes at least one device configured to actuate one or more brakes to decelerate the vehicle 200 and / or keep it stationary. In some examples, the braking system 208 includes at least one controller and / or actuator configured to close one or more calipers associated with one or more wheels of the vehicle 200 on the corresponding rotors of the vehicle 200. Additionally or alternatively, in some examples, the braking system 208 includes an automatic emergency braking (AEB) system and / or a regenerative braking system, etc.

[0062] In some embodiments, the vehicle 200 includes at least one platform sensor (not explicitly illustrated) for measuring or inferring the nature of the state or condition of the vehicle 200. In some examples, the vehicle 200 includes platform sensors such as a global positioning system (GPS) receiver, an inertial measurement unit (IMU), a wheel speed sensor, a wheel brake pressure sensor, a wheel torque sensor, an engine torque sensor, and / or a steering angle sensor. Although the braking system 208 is illustrated as being located Figure 2 proximal to the vehicle 200 within, the braking system 208 can be located anywhere within the vehicle 200.

[0063] Now refer to Figure 3 , a schematic diagram of the illustrative device 300. As illustrated, the device 300 includes a processor 304, a memory 306, a storage component 308, an input interface 310, an output interface 312, a communication interface 314, and a bus 302. In some embodiments, the device 300 corresponds to: at least one device of the vehicle 102 (e.g., at least one device of the system of the vehicle 102); and / or one or more devices of the network 112 (e.g., one or more devices of the system of the network 112). In some embodiments, one or more devices of the vehicle 102 (e.g., one or more devices of the system of the vehicle 102), and / or one or more devices of the network 112 (e.g., one or more devices of the system of the network 112) include at least one device 300 and / or at least one component of the device 300. As Figure 3As shown, device 300 includes bus 302, processor 304, memory 306, storage component 308, input interface 310, output interface 312, and communication interface 314.

[0064] Bus 302 includes components that permit communication between the components of device 300. In some cases, processor 304 includes a processor (e.g., a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), and / or a neural processing unit (NPU), etc.), a digital signal processor (DSP), and / or any processing component that can be programmed to perform at least one function (e.g., a field programmable gate array (FPGA) and / or an application specific integrated circuit (ASIC), etc.). Memory 306 includes random access memory (RAM), read only memory (ROM), and / or another type of dynamic and / or static storage device that stores data and / or instructions for use by processor 304 (e.g., flash memory, magnetic memory, optical memory, and / or dynamic RAM (DRAM), etc.).

[0065] Storage component 308 stores data and / or software related to the operation and use of device 300. In some examples, storage component 308 includes a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optical disk, and / or a solid state disk, etc.), a compact disk (CD), a digital versatile disk (DVD), a floppy disk, a cassette tape, a magnetic tape, a CD-ROM, RAM, PROM, EPROM, FLASH-EPROM, NV-RAM, and / or another type of computer readable medium, and a corresponding drive.

[0066] Input interface 310 includes components that permit device 300 to receive information such as via user input (e.g., a touch screen display, a keyboard, a keypad, a mouse, a button, a switch, a microphone, and / or a camera, etc.). Additionally or alternatively, in some embodiments, input interface 310 includes sensors for sensing information (e.g., a global positioning system (GPS) receiver, an accelerometer, a gyroscope, and / or an actuator, etc.). Output interface 312 includes components for providing output information from device 300 (e.g., a display, a speaker, and / or one or more light emitting diodes (LEDs), etc.).

[0067] In some embodiments, communication interface 314 includes transceiver-like components (e.g., transceivers and / or separate receivers and transmitters, etc.) that permit device 300 to communicate with other devices via a wired connection, a wireless connection, or a combination of a wired connection and a wireless connection. In some examples, communication interface 314 permits device 300 to receive information from and / or provide information to another device. In some examples, communication interface 314 includes an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, an interface and / or a cellular network interface, etc.

[0068] In some embodiments, device 300 performs one or more processes described herein. Device 300 performs these processes based on software instructions stored by a computer-readable medium such as memory 306 and / or storage component 308 and executed by processor 304. A computer-readable medium (e.g., a non-transitory computer-readable medium) is defined herein as a non-transitory memory device. A non-transitory memory device includes a storage space located within a single physical storage device or a storage space distributed across multiple physical storage devices.

[0069] In some embodiments, software instructions are read into memory 306 and / or storage component 308 from another computer-readable medium or from another device via communication interface 314. When executed, the software instructions stored in memory 306 and / or storage component 308 cause processor 304 to perform one or more processes described herein. Additionally or alternatively, hardwired circuitry is used instead of or in combination with software instructions to perform one or more processes described herein. Accordingly, unless otherwise expressly stated, the embodiments described herein are not limited to any particular combination of hardware circuitry and software.

[0070] Memory 306 and / or storage component 308 includes a data store or at least one data structure (e.g., a database, etc.). Device 300 is capable of receiving information from, storing information in, communicating information to, or searching for information stored in the data store or at least one data structure in memory 306 or storage component 308. In some examples, the information includes network data, input data, output data, or any combination thereof.

[0071] In some embodiments, apparatus 300 is configured to execute software instructions stored in memory 306 and / or the memory of another apparatus (e.g., another apparatus that is the same as or similar to apparatus 300). As used herein, the term "module" refers to at least one instruction stored in memory 306 and / or the memory of another apparatus, which, when executed by processor 304 and / or the processor of another apparatus (e.g., another apparatus that is the same as or similar to apparatus 300), causes apparatus 300 (e.g., at least one component of apparatus 300) to perform one or more processes described herein. In some embodiments, the module is implemented in software, firmware, and / or hardware, etc.

[0072] Provide Figure 3 The number and arrangement of the illustrated components are provided as examples. In some embodiments, compared to Figure 3 the illustrated components, apparatus 300 may include additional components, fewer components, different components, or components in a different arrangement. Additionally or alternatively, a set of components of apparatus 300 (e.g., one or more components) may perform one or more functions described as being performed by another component or another set of components of apparatus 300.

[0073] Now refer to Figure 4A, an example block diagram of autonomous vehicle software 400 (sometimes referred to as the "AV stack") is illustrated. As illustrated, autonomous vehicle software 400 includes a perception system 402 (sometimes referred to as the perception module), a planning system 404 (sometimes referred to as the planning module), a localization system 406 (sometimes referred to as the localization module), a control system 408 (sometimes referred to as the control module), and a database 410. In some embodiments, the perception system 402, the planning system 404, the localization system 406, the control system 408, and the database 410 are included in and / or implemented in the vehicle's automatic navigation system (e.g., the autonomous vehicle computing 202f of vehicle 200). Additionally or alternatively, in some embodiments, the perception system 402, the planning system 404, the localization system 406, the control system 408, and the database 410 are included in one or more standalone systems (e.g., one or more systems that are the same as or similar to the autonomous vehicle software 400, etc.). In some examples, the perception system 402, the planning system 404, the localization system 406, the control system 408, and the database 410 are included in one or more standalone systems located in the vehicle and / or at least one remote system as described herein. In some embodiments, any and / or all of the systems included in the autonomous vehicle software 400 are implemented by computer hardware (e.g., via a microprocessor, a microcontroller, an application specific integrated circuit (ASIC), and / or a field programmable gate array (FPGA), etc.), a chiplet, or a distributed computing architecture with software (e.g., software instructions stored in memory). It will also be understood that, in some embodiments, the autonomous vehicle software 400 is configured to communicate with remote systems (e.g., an autonomous vehicle system that is the same as or similar to the remote AV system 114, a queue management system that is the same as or similar to the queue management system 116, and / or a V2I system that is the same as or similar to the V2I system 118, etc.).

[0074] In some embodiments, the perception system 402 receives data associated with at least one physical object in the environment (e.g., data used by the perception system 402 to detect at least one physical object), and classifies the at least one physical object. In some examples, the perception system 402 receives image data captured by at least one camera (e.g., camera 202a), the image being associated with one or more physical objects within the field of view of the at least one camera (e.g., representing the one or more physical objects). In such examples, the perception system 402 classifies the at least one physical object based on one or more groupings of physical objects (e.g., bicycles, vehicles, traffic signs, and / or pedestrians, etc.). In some embodiments, based on the classification of the physical object by the perception system 402, the perception system 402 transmits data associated with the classification of the physical object to the planning system 404.

[0075] In some embodiments, the planning system 404 receives data associated with a destination and generates data associated with at least one route (e.g., route 106) along which a vehicle (e.g., vehicle 102) can travel toward the destination. In some embodiments, the planning system 404 receives data (e.g., the data associated with the classification of physical objects described above) from the sensing system 402 periodically or continuously, and the planning system 404 updates at least one trajectory or generates at least one different trajectory based on the data generated by the sensing system 402. In other words, the planning system 404 can perform tasks related to tactical functions required to operate the vehicle 102 in road traffic. Tactical efforts involve maneuvering the vehicle in traffic during a journey, which includes but is not limited to deciding whether and when to overtake another vehicle, change lanes, or select an appropriate speed, acceleration, deceleration, etc. In some embodiments, the planning system 404 receives data associated with the updated position of a vehicle (e.g., vehicle 102) from the positioning system 406, and the planning system 404 updates at least one trajectory or generates at least one different trajectory based on the data generated by the positioning system 406.

[0076] In some embodiments, the positioning system 406 receives data associated with (e.g., representing) the location of a vehicle (e.g., vehicle 102) in an area. In some examples, the positioning system 406 receives LiDAR data associated with at least one point cloud generated by at least one LiDAR sensor (e.g., LiDAR sensor 202b). In certain examples, the positioning system 406 receives data associated with at least one point cloud from multiple LiDAR sensors, and the positioning system 406 generates a combined point cloud based on the respective point clouds. In these examples, the positioning system 406 compares the at least one point cloud or the combined point cloud with two-dimensional (2D) and / or three-dimensional (3D) maps of the area stored in the database 410. Then, based on the positioning system 406 comparing the at least one point cloud or the combined point cloud with the map, the positioning system 406 determines the position of the vehicle in the area. In some embodiments, the map includes a combined point cloud of the area generated prior to navigation of the vehicle. In some embodiments, the map includes, but is not limited to, a high-precision map of the roadway geometry, a map describing the connectivity of the road network, a map describing the physical properties of the roadways (such as traffic speed, traffic flow, the number of vehicle and bicycle traffic lanes, lane width, lane traffic direction or the type and location of lane markings, or a combination thereof, etc.), and a map describing the spatial location of road features (such as crosswalks, traffic signs, or other types of driving signals of various types, etc.). In some embodiments, the map is generated in real time based on the data received by the sensing system.

[0077] In another example, the positioning system 406 receives global navigation satellite system (GNSS) data generated by a global positioning system (GPS) receiver. In some examples, the positioning system 406 receives GNSS data associated with the location of the vehicle in the area, and the positioning system 406 determines the latitude and longitude of the vehicle in the area. In such examples, the positioning system 406 determines the position of the vehicle in the area based on the latitude and longitude of the vehicle. In some embodiments, the positioning system 406 generates data associated with the position of the vehicle. In some examples, based on the positioning system 406 determining the position of the vehicle, the positioning system 406 generates data associated with the position of the vehicle. In such examples, the data associated with the position of the vehicle includes data associated with one or more semantic properties corresponding to the position of the vehicle.

[0078] In some embodiments, the control system 408 receives data associated with at least one trajectory from the planning system 404, and the control system 408 controls the operation of the vehicle. In some examples, the control system 408 receives data associated with at least one trajectory from the planning system 404, and the control system 408 controls the operation of the vehicle by generating and transmitting control signals to cause the powertrain control system (e.g., DBW system 202h and / or powertrain control system 204, etc.), the steering control system (e.g., steering control system 206), and / or the braking system (e.g., braking system 208) to operate. For example, the control system 408 is configured to perform operational functions such as lateral vehicle motion control or longitudinal vehicle motion control. Lateral vehicle motion control causes activities required to regulate the y-axis component of the vehicle motion. Longitudinal vehicle motion control causes activities required to regulate the x-axis component of the vehicle motion. In an example, in the case where the trajectory includes a left turn, the control system 408 transmits a control signal to cause the steering control system 206 to adjust the steering angle of the vehicle 200, thereby causing the vehicle 200 to turn left. Additionally or alternatively, the control system 408 generates and transmits control signals to cause other devices of the vehicle 200 (e.g., headlights, turn signals, door locks, and / or windshield wipers, etc.) to change states.

[0079] In some embodiments, the perception system 402, the planning system 404, the positioning system 406, and / or the control system 408 implement at least one machine learning model (e.g., at least one multi-layer perceptron (MLP), at least one convolutional neural network (CNN), at least one recurrent neural network (RNN), at least one autoencoder, and / or at least one transformer, etc.). In some examples, the perception system 402, the planning system 404, the positioning system 406, and / or the control system 408 implement at least one machine learning model alone or in combination with one or more of the above systems. In some examples, the perception system 402, the planning system 404, the positioning system 406, and / or the control system 408 implement at least one machine learning model as part of a pipeline (e.g., a pipeline for identifying one or more objects located in the environment, etc.). The following are examples of Figures 4B to 4D implementations that include machine learning models.

[0080] The database 410 stores data transmitted to, received from, and / or updated by the perception system 402, the planning system 404, the positioning system 406, and / or the control system 408. In some examples, the database 410 includes a storage component of at least one system for storing operation-related data and / or software and using the autonomous vehicle software 400 (e.g., a storage component Figure 3 similar to or the same as the storage component 308). In some embodiments, the database 410 stores data associated with 2D and / or 3D maps of at least one area. In some examples, the database 410 stores data associated with 2D and / or 3D maps of a part of a city, multiple parts of multiple cities, multiple cities, counties, states, and / or countries (e.g., a country). In such examples, a vehicle (e.g., a vehicle similar to or the same as the vehicle 102 and / or the vehicle 200) can drive along one or more drivable areas (e.g., single-lane roads, multi-lane roads, highways, unpaved roads, and / or off-road roads, etc.), and cause at least one LiDAR sensor (e.g., a LiDAR sensor similar to or the same as the LiDAR sensor 202b) to generate data associated with an image representing the objects included in the field of view of the at least one LiDAR sensor.

[0081] In some embodiments, the database 410 can be implemented across multiple devices. In some examples, the database 410 includes components in a vehicle (e.g., a vehicle similar to or the same as the vehicle 102 and / or the vehicle 200), an autonomous vehicle system (e.g., an autonomous vehicle system similar to or the same as the remote AV system 114), a queue management system (e.g., similar to Figure 1in the same or similar queue management system 116) and / or in a V2I system (e.g., the same or similar V2I system 118 as the V2I system). Figure 1 in the V2I system).

[0082] Now refer to Figure 4B , a diagram illustrating the implementation of a machine learning model. More specifically, a diagram illustrating the implementation of a convolutional neural network (CNN) 420. For illustrative purposes, the following description of the CNN 420 will be with respect to the implementation of the CNN 420 by the sensing system 402. However, it will be understood that in some examples, the CNN 420 (e.g., one or more components of the CNN 420) is implemented by other systems different from or in addition to the sensing system 402, such as the planning system 404, the positioning system 406, and / or the control system 408, etc. Although the CNN 420 includes certain features as described herein, these features are provided for illustrative purposes and are not intended to limit the present disclosure.

[0083] The CNN 420 includes a plurality of convolutional layers including a first convolutional layer 422, a second convolutional layer 424, and a convolutional layer 426. In some embodiments, the CNN 420 includes a subsampling layer 428 (sometimes referred to as a pooling layer). In some embodiments, the subsampling layer 428 and / or other subsampling layers have a dimension smaller than the dimension of the upstream system (i.e., the amount of nodes). By means of the subsampling layer 428 having a dimension smaller than the dimension of the upstream layer, the CNN 420 combines the amount of data associated with the initial input and / or output of the upstream layer, thereby reducing the amount of computation required for the CNN 420 to perform downstream convolutional operations. Additionally or alternatively, by means of the subsampling layer 428 being associated with at least one subsampling function (e.g., being configured to perform at least one subsampling function) (as described below with respect to Figure 4C and Figure 4D ), the CNN 420 combines the amount of data associated with the initial input.

[0084] Based on the perception system 402 providing corresponding inputs and / or outputs associated with the first convolutional layer 422, the second convolutional layer 424, and the convolutional layer 426 respectively to generate corresponding outputs, the perception system 402 performs convolutional operations. In some examples, based on the perception system 402 providing data as inputs to the first convolutional layer 422, the second convolutional layer 424, and the convolutional layer 426, the perception system 402 implements the CNN 420. In such examples, based on the perception system 402 receiving data from one or more different systems (e.g., one or more systems of a vehicle identical or similar to the vehicle 102, a remote AV system identical or similar to the remote AV system 114, a queue management system identical or similar to the queue management system 116, and / or a V2I system identical or similar to the V2I system 118, etc.), the perception system 402 provides the data as inputs to the first convolutional layer 422, the second convolutional layer 424, and the convolutional layer 426. The following is a detailed description of Figure 4C the convolutional operations.

[0085] In some embodiments, the perception system 402 provides data associated with an input (referred to as an initial input) to the first convolutional layer 422, and the perception system 402 uses the first convolutional layer 422 to generate data associated with an output. In some embodiments, the perception system 402 provides the output generated by a convolutional layer as an input to a different convolutional layer. For example, the perception system 402 provides the output of the first convolutional layer 422 as an input to the subsampling layer 428, the second convolutional layer 424, and / or the convolutional layer 426. In such examples, the first convolutional layer 422 is referred to as an upstream layer, and the subsampling layer 428, the second convolutional layer 424, and / or the convolutional layer 426 are referred to as downstream layers. Similarly, in some embodiments, the perception system 402 provides the output of the subsampling layer 428 to the second convolutional layer 424 and / or the convolutional layer 426, and in this example, the subsampling layer 428 will be referred to as an upstream layer, and the second convolutional layer 424 and / or the convolutional layer 426 will be referred to as downstream layers.

[0086] In some embodiments, before the perception system 402 provides an input to the CNN 420, the perception system 402 processes the data associated with the input provided to the CNN 420. For example, based on the perception system 402 normalizing sensor data (e.g., image data, LiDAR data, and / or Radar data, etc.), the perception system 402 processes the data associated with the input provided to the CNN 420.

[0087] In some embodiments, convolution operations associated with respective convolutional layers are performed based on the perception system 402, and the CNN 420 generates an output. In some examples, convolution operations associated with respective convolutional layers and the initial input are performed based on the perception system 402, and the CNN 420 generates an output. In some embodiments, the perception system 402 generates an output and provides the output to the fully connected layer 430. In some examples, the perception system 402 provides the output of the convolutional layer 426 to the fully connected layer 430, where the fully connected layer 430 includes data associated with a plurality of eigenvalues referred to as F1, F2, ..., FN. In this example, the output of the convolutional layer 426 includes data associated with a plurality of output eigenvalues representing predictions.

[0088] In some embodiments, the perception system 402 identifies the eigenvalue associated with the highest likelihood of being the correct prediction among a plurality of predictions, and the perception system 402 identifies a prediction from among the plurality of predictions. For example, in a case where the fully connected layer 430 includes eigenvalues F1, F2, ..., FN and F1 is the largest eigenvalue, the perception system 402 identifies the prediction associated with F1 as the correct prediction among the plurality of predictions. In some embodiments, the perception system 402 trains the CNN 420 to generate predictions. In some examples, based on the perception system 402 providing training data associated with the predictions to the CNN 420, the perception system 402 trains the CNN 420 to generate predictions.

[0089] Now refer to Figure 4C and Figure 4D , a diagram illustrating an example operation of the CNN 440 that utilizes the perception system 402. In some embodiments, the CNN 440 (e.g., one or more components of the CNN 440) is the same as or similar to the CNN 420 (e.g., one or more components of the CNN 420) (see Figure 4B ).

[0090] In step 450, the perception system 402 provides data associated with an image as an input to the CNN 440 (step 450). For example, as illustrated, the perception system 402 provides data associated with an image to the CNN 440, where the image is a grayscale image represented as values stored in a two-dimensional (2D) array. In some embodiments, the data associated with the image may include data associated with a color image, which is represented as values stored in a three-dimensional (3D) array. Additionally or alternatively, the data associated with the image may include data associated with an infrared image and / or a Radar image, etc.

[0091] At step 455, CNN 440 performs a first convolution function. For example, based on CNN 440 providing values representing an image as input to one or more neurons (not explicitly illustrated) included in the first convolutional layer 442, CNN 440 performs the first convolution function. In this example, the values representing the image may correspond to values of a region (sometimes referred to as a receptive field) representing the image. In some embodiments, each neuron is associated with a filter (not explicitly illustrated). The filter (sometimes referred to as a kernel) can be represented as an array of values corresponding in size to the values provided as input to the neuron. In one example, the filter can be configured to identify edges (e.g., horizontal lines, vertical lines, and / or straight lines, etc.). In successive convolutional layers, the filters associated with the neurons can be configured to successively identify more complex patterns (e.g., arcs and / or objects, etc.).

[0092] In some embodiments, based on CNN 440 multiplying the values provided as input to each neuron of one or more neurons included in the first convolutional layer 442 by the values of the filters corresponding to each neuron of the same or more neurons, CNN 440 performs the first convolution function. For example, CNN 440 can multiply the values provided as input to each neuron of one or more neurons included in the first convolutional layer 442 by the values of the filters corresponding to each neuron of the one or more neurons to generate a single value or an array of values as output. In some embodiments, the collective output of the neurons of the first convolutional layer 442 is referred to as the convolution output. In some embodiments, when each neuron has the same filter, the convolution output is referred to as a feature map.

[0093] In some embodiments, CNN 440 provides the output of each neuron of the first convolutional layer 442 to neurons of a downstream layer. For clarity, an upstream layer can be a layer that transmits data to a different layer (referred to as a downstream layer). For example, CNN 440 can provide the output of each neuron of the first convolutional layer 442 to the corresponding neurons of a subsampling layer. In the example, CNN 440 provides the output of each neuron of the first convolutional layer 442 to the corresponding neurons of the first subsampling layer 444. In some embodiments, CNN 440 adds a bias value to the aggregate set of all values provided to each neuron of the downstream layer. For example, CNN 440 adds a bias value to the aggregate set of all values provided to each neuron of the first subsampling layer 444. In such an example, CNN 440 determines the final value to be provided to each neuron of the first subsampling layer 444 based on the aggregate set of all values provided to each neuron and an activation function associated with each neuron of the first subsampling layer 444.

[0094] At step 460, the CNN 440 performs a first subsampling function. For example, based on the values output by the first convolutional layer 442 being provided to the respective neurons of the first subsampling layer 444, the CNN 440 may perform the first subsampling function. In some embodiments, the CNN 440 performs the first subsampling function based on an aggregation function. In an example, based on the CNN 440 determining the maximum input among the values provided to a given neuron (referred to as the max pooling function), the CNN 440 performs the first subsampling function. In another example, based on the CNN 440 determining the average input among the values provided to a given neuron (referred to as the average pooling function), the CNN 440 performs the first subsampling function. In some embodiments, based on the CNN 440 providing values to the respective neurons of the first subsampling layer 444, the CNN 440 generates an output, which is sometimes referred to as the subsampled convolutional output.

[0095] At step 465, the CNN 440 performs a second convolutional function. In some embodiments, the CNN 440 performs the second convolutional function in a manner similar to how the CNN 440 performs the first convolutional function as described above. In some embodiments, based on the values output by the first subsampling layer 444 being provided as inputs to one or more neurons (not explicitly illustrated) included in the second convolutional layer 446, the CNN 440 performs the second convolutional function. In some embodiments, as described above, each neuron of the second convolutional layer 446 is associated with a filter. As described above, the (one or more) filters associated with the second convolutional layer 446 may be configured to identify more complex patterns compared to the filters associated with the first convolutional layer 442.

[0096] In some embodiments, based on the CNN 440 multiplying the values provided as inputs to each of the one or more neurons included in the second convolutional layer 446 by the values of the filters corresponding to each of the one or more neurons, the CNN 440 performs the second convolutional function. For example, the CNN 440 may multiply the values provided as inputs to each of the one or more neurons included in the second convolutional layer 446 by the values of the filters corresponding to each of the one or more neurons to generate a single value or an array of values as an output.

[0097] In some embodiments, the CNN 440 provides the outputs of the respective neurons of the second convolutional layer 446 to the neurons of the downstream layer. For example, the CNN 440 may provide the outputs of the respective neurons of the first convolutional layer 442 to the corresponding neurons of the subsampling layer. In an example, the CNN 440 provides the outputs of the respective neurons of the first convolutional layer 442 to the corresponding neurons of the second subsampling layer 448. In some embodiments, the CNN 440 adds a bias value to the aggregated set of all values provided to each neuron of the downstream layer. For example, the CNN 440 adds a bias value to the aggregated set of all values provided to each neuron of the second subsampling layer 448. In such an example, the CNN 440 determines the final value provided to each neuron of the second subsampling layer 448 based on the aggregated set of all values provided to each neuron and the activation function associated with each neuron of the second subsampling layer 448.

[0098] In step 470, the CNN 440 performs a second subsampling function. For example, based on the CNN 440 providing the values output by the second convolutional layer 446 to the corresponding neurons of the second subsampling layer 448, the CNN 440 may perform a second subsampling function. In some embodiments, based on the CNN 440 using an aggregation function, the CNN 440 performs a second subsampling function. In an example, as described above, based on the CNN 440 determining the maximum input or average input among the values provided to a given neuron, the CNN 440 performs a first subsampling function. In some embodiments, based on the CNN 440 providing values to each neuron of the second subsampling layer 448, the CNN 440 generates an output.

[0099] In step 475, the CNN 440 provides the outputs of the respective neurons of the second subsampling layer 448 to the fully connected layer 449. For example, the CNN 440 provides the outputs of the respective neurons of the second subsampling layer 448 to the fully connected layer 449 such that the fully connected layer 449 generates an output. In some embodiments, the fully connected layer 449 is configured to generate an output associated with a prediction (sometimes referred to as classification). The prediction may include an indication that the objects included in the image provided as input to the CNN 440 include objects and / or a set of objects, etc. In some embodiments, the perception system 402 performs one or more operations and / or provides data associated with the prediction to different systems described herein.

[0100] Figure 5 is a block diagram of the DETR encoder-decoder architecture 500. The architecture 500 includes a backbone 501, a transformer encoder 502, a transformer decoder 503, and a prediction head 504. The Figure 5 number and arrangement of the illustrated components are provided as examples. In some embodiments, Figure 5Compared with the illustrated components, the architecture 500 may include additional components, fewer components, different components, or components arranged differently. Additionally or alternatively, a set of components (e.g., one or more than one component) of the architecture 500 may perform one or more than one function described as being performed by another component or another set of components of the architecture 500.

[0101] For illustrative purposes, the following description of the architecture 500 (referred to as the DETR architecture 500) will be directed to the implementation of the architecture 500 by the perception system 402. However, it will be understood that in some examples, the DETR architecture 500 (e.g., one or more than one component of the DETR architecture 500) is implemented by other systems different from or in addition to the perception system 402, such as the planning system 404, the positioning system 406, and / or the control system 408, etc. Although the DETR architecture 500 includes certain features as described herein, these features are provided for illustrative purposes and are not intended to limit the present disclosure. Additionally, the DETR architecture 500 may be replaced with any end-to-end detector, including but not limited to: deformable DETR, sparse RCNN, and any other deep learning architecture capable of object detection. The DETR architecture 500 may be implemented within any deep learning architecture for object detection and tracking.

[0102] The backbone 501 includes a CNN (e.g., CNN 440) to learn two-dimensional (2D) lower-resolution feature maps of the input image (e.g., having 3 color channels). The transformer encoder 502 performs 1×1 convolutions on this feature map to reduce the channel dimension to a smaller dimension, thereby creating a new feature map. Each layer of the transformer encoder 502 includes a multi-head self-attention module and a feed-forward network (FFN). Since the transformer encoder 502 expects a sequence as input, the transformer encoder 502 flattens the spatial dimension of this new feature map into one dimension. Since the DETR architecture 500 is permutation-invariant, a fixed positional encoding 506 is added (e.g., concatenated) to the input to each attention layer in the transformer encoder 502.

[0103] The transformer decoder 503 includes a multi-head self-attention mechanism and an encoder-decoder attention mechanism (not explicitly illustrated). The transformer decoder 503 takes a small fixed number of learned positional embeddings (referred to as object queries) as input and adds them to the input of each attention layer. Each output embedding of the transformer decoder 503 is passed to the prediction head 504, which includes a shared feed-forward network (FFN) for predicting (respectively) classes / bounding boxes or predicting the "no object" class.

[0104] In some embodiments, the FNN includes a 3-layer perceptron with ReLU activation function and hidden dimension d, and a linear projection layer. The FFN predicts the normalized center coordinates, height, and width of the bounding boxes relative to the input image, and the linear layer uses the softmax function to predict the class labels. Since a fixed-size set of N bounding boxes is predicted (where N is typically much larger than the actual number of objects of interest in the image), an "objectness" class is used to indicate that no object is detected in a slot. The "objectness" class plays a role similar to the "background" class in standard object detection methods.

[0105] The parameters (e.g., weights) of the above-described DETR architecture 500 can be trained using backpropagation techniques that minimize a global loss function as described in the seminal paper. However, as described below, the training can be improved in terms of speed and accuracy by representing the global loss function as the sum of an assignment-independent term and an assignment-dependent term.

[0106] After the DETR architecture 500 is trained, it can be used in various applications, including but not limited to being used in the perception system 402 of the AV stack of an AV to detect and track objects in an AV operating environment.

[0107] New representation for the global loss

[0108] In the following disclosure, some notations from the seminal paper are adopted, but new notations are also defined. Let be the set of M ground-truth objects, and be the set of N predictions. At inference time, thresholding the probabilities or scores will give a subset of the N predictions as the actual objects. At training time, it is desired to accurately select M predictions from the N predictions to correspond to the ground-truth annotations. Since typically N >> M, it is desired to solve the assignment problem using a rectangular cost matrix rather than a square cost matrix (as done in the code released by the authors of the seminal paper). Thus, instead of talking about permutations, an injective mapping s: G → B from the set of ground-truth indices to the set of predicted box indices is defined. The matching set is defined as and the non-matching set is defined as Naturally defines the inverse mapping s -1 : B1 → G. In other words, starting from the ground-truth annotated by index j, the assigned prediction is indexed by s(j). Starting from the prediction indexed by i and known to be assigned a ground-truth annotation, the ground-truth index is s -1 (i).

[0109] As described in the seminal paper, let y i =(c j , bj ) is the ground truth annotation with object class label c j and bounding box vector b j . Let be the loss when the predicted bounding box is considered to represent b j . Let represent the probability of the target object class in the i-th prediction, and represent the probability of the background

[0110] For a given set of network weight vectors w and a given mapping s, the loss is related to three parts: 1) the object class probabilities in the matching set B1; 2) the background probabilities in the "rest" set B2; and 3) the loss between G and B1 The third part is the same as described in the pioneering paper, and the details will be omitted here. The first two parts depend on the following likelihoods

[0111]

[0112] Note that the two terms in Equation [1] depend on the mapping s. However, the following terms do not depend on the mapping s

[0113]

[0114] Thus, the following ratio

[0115]

[0116] depends only on the index sets G and B1. Therefore, the global loss (which is called the Hungarian algorithm loss in the pioneering paper) is written as follows

[0117]

[0118] The new cost matrix

[0119] The matching-related term in Equation [4] can be written as the total assignment cost

[0120]

[0121] where the (i, j)-th entry of the N×M cost matrix C is defined as

[0122]

[0123] Now, it is clear that minimizing the global loss involves the following two steps

[0124] 1. Solve the assignment problem using the cost matrix defined in Equation [6] as follows

[0125]

[0126] 2. Minimize the following:

[0127]

[0128] Generalized gradients

[0129] The optimal assignment cost can be obtained using the solution given in Xi Gao, Han Zhang, Aliakbar Panahi, and Tom Arodz. Combinatorial losses through generalized gradients of integer linear programs. arXiv preprint arXiv:1910.08211, 2019 The generalized gradients with respect to the network weights w. Thus, as described in Frank H Clarke. Optimization and nonsmooth analysis. SIAM, 1990, the generalized gradients of can be obtained and used in training.

[0130] More specifically, the cost matrix C in Equation [6] is defined by the ground truth label y and the network prediction as a function of the network weight vector w If the columns of C are stacked to form the cost vector c(w), then as described in David F Crouse. On implementing 2d rectangular assignment algorithms. IEEE Transactions on Aerospace and Electronic Systems, 52:1679 - 1696, 2016, the assignment problem can be formulated as an ILP problem with parameters (c(w), A, b), where both A and b are constants that specify the inequality constraints for a valid assignment. The cost matrix can be filled with random values that are nearly zero to form a square matrix with equality constraints so that the results can be applied as in Xi Gao, Han Zhang, Aliakbar Panahi, and Tom Arodz. Combinatorial losses through generalized gradients of integer linear programs. arXiv preprint arXiv:1910.08211, 2019

[0131] Let is the optimal solution to the assignment problem (in the form of a column vector); in fact, it is almost always unique. Then, according to Algorithm 1 in Xi Gao, Han Zhang, Aliakbar Panahi, and Tom Arodz. Combinatorial losses through generalized gradients of integer linear programs. arXiv preprint arXiv:1910.08211, 2019, we have the following:

[0132]

[0133] This provides a way to compute gradients using PyTorch by treating the optimal assignment solution as a constant. This means taking only the numerical values of the cost matrix and (under torch.no_grad() in the code based on the seminal paper) calling the Hungarian algorithm solver. Based on the solution sum the optimal cost attached to the gradient together with the background term, and backpropagate the loss in Equation [8].

[0134] Figure 6 is a flowchart of the processing 600 by the DETR transformer according to one or more embodiments. In some embodiments, the processing 600 can be performed by the sensing system 402.

[0135] In some embodiments, the processing 600 includes: initializing the parameters of the DETR (601); propagating the input image through the DETR (602); determining an error value by comparing the output of the DETR with a known expected output (603); and iteratively updating the parameters in the DETR based on the error value by minimizing the global loss. In some embodiments, minimizing the global loss includes: generating an assignment cost matrix (604); solving an assignment optimization problem based on the assignment cost matrix to obtain an optimal assignment cost (605); and then minimizing a global loss function represented as the sum of the optimal assignment cost and the assignment-independent loss (606).

[0136] The above-mentioned process 600 has the following advantages compared to the process disclosed in the pioneering paper: 1) The allocation cost and the global loss are consistent; 2) The unmatched predictions have different background probabilities in themselves depending on the allocation, so they are not ignored in the cost matrix; 3) Instead of the original probability, the cross-entropy loss is used in the new cost matrix (if the scaling problem still exists, this problem can be solved separately); 4) The challenge of the rectangular cost matrix is solved by using the ratio in Equation [3] that gives the cost function in Equation [6], that is, subtracting the cross-entropy loss corresponding to the background from each row; 5) There is a normalization by the number of boxes in the definition of the global loss for the loss based on the Generalized Intersection over Union (GIOU), while it is not done in the total allocation cost. The loss represented by Equation [4] cannot justify the normalization, that is, batches with more boxes have a greater impact.

[0137] In the foregoing description, aspects and embodiments of the present disclosure have been described with reference to numerous specific details, which may vary depending on the implementation. Accordingly, the specification and drawings are to be regarded as illustrative rather than restrictive in a limiting sense. The sole and exclusive indication of the scope of the invention, and what the applicant desires to be the scope of the invention, is the literal and equivalent scope of the claims as issued from this application in the specific form of the claims as issued in the patent grant, including any subsequent amendments. Any definition of terms expressly set forth herein for inclusion in such claims shall be construed to have the meaning such terms have as used in the claims. Additionally, when the term "further comprises" is used in the foregoing specification or the appended claims, the text following this phrase may be an additional step or entity, or a sub-step / sub-entity of a previously described step or entity.

Claims

1. A method for training a detection transformer, i.e., DETR, includes: Initializing the parameters of the DETR using at least one processor; Propagating an input image through the DETR using the at least one processor; Determining an error value by comparing the output of the DETR with a known expected output using the at least one processor; And Iteratively updating the parameters in the DETR based on the error value by minimizing a global loss using the at least one processor, where minimizing the global loss includes: Generating an assignment cost matrix; Solving an assignment optimization problem based on the assignment cost matrix to obtain an optimal assignment cost; and Minimizing a global loss function represented as the sum of the optimal assignment cost and an assignment-independent loss.

2. The method according to claim 1, wherein, The assignment cost matrix is a rectangular matrix.

3. The method according to claim 1, wherein, The assignment cost matrix is defined by ground truth annotations and network predictions as a function of network weight operators.

4. The method according to claim 1, further comprising: Filling the assignment cost matrix with random values that are nearly zero to form a square matrix.

5. The method according to claim 1 further comprises: Subtracting the cross-entropy loss corresponding to the background class from each row of the assignment cost matrix.

6. The method according to claim 1 further comprises: Using a Hungarian algorithm solver to solve the assignment problem.

7. The method according to claim 1, wherein, The assignment-independent term is an unmatched prediction loss.

8. An object detector includes: A convolutional layer; A transformer encoder; A transformer decoder; And At least one prediction head, where the parameters of the detection transformer are determined by: Propagating an input image through the object detector; Determining an error value by comparing the output of the at least one prediction head with a known expected output; and Iteratively updating the parameters in the object detector based on the error value by minimizing a global loss, where minimizing the global loss includes: generating an assignment cost matrix; using the assignment cost matrix to solve an assignment optimization problem to obtain an optimal assignment cost; and then minimizing a global loss function represented as the sum of the optimal assignment cost and an assignment-independent loss.

9. The object detector according to claim 8, wherein, The convolutional layer is a convolutional neural network.

10. The object detector according to claim 9, wherein, The assignment cost matrix is a rectangular matrix.

11. The object detector according to claim 8, wherein, The assignment cost matrix is defined by ground truth annotations and network predictions as a function of network weight operators.

12. The object detector according to claim 8, wherein, Filling the assignment cost matrix with random values that are nearly zero to form a square matrix.

13. The object detector according to claim 8, wherein, Subtracting the cross-entropy loss corresponding to the background class from each row of the assignment cost matrix.

14. The object detector according to claim 8, wherein, Using a Hungarian algorithm solver to solve the assignment problem.

15. The object detector according to claim 8, wherein, The assignment-independent loss is an unmatched prediction loss.

16. A method includes: Obtaining image data associated with an image from at least one sensor of a vehicle using at least one processor, the image being captured by the at least one sensor; Detecting at least one object captured in the image using the at least one processor, where the detection includes: Input the image data into a detection transformer, i.e., DETR, which has parameters determined by minimizing a global loss during a training process, where minimizing the global loss includes: generating an assignment cost matrix; solving an assignment optimization problem based on the assignment cost matrix to obtain an optimal assignment cost; and then minimizing a global loss function, which is expressed as the sum of the optimal assignment cost and an assignment-independent loss; and Using the at least one processor, generate at least one control signal for controlling the vehicle, at least in part based on the at least one detected object.

17. The method according to claim 16, wherein, The assignment-independent loss is an unmatched prediction loss.

18. The method according to claim 16, wherein, The assignment cost matrix is defined by ground truth annotations and network predictions as a function of network weight operators.

19. The method according to claim 16 further comprises: Fill the assignment cost matrix with random values that are nearly zero to form a square matrix.

20. The method according to claim 16, further comprising: Subtract the cross-entropy loss corresponding to the background from each row of the assignment cost matrix.

Images