Automatically select data samples for annotation

Abstract

Techniques for automatically selecting data samples for annotation are described, among other things. The techniques select data samples for annotation using bounding box predictions based on bounding box score distributions, spatial probability densities determined from bounding box sizes and locations, and overall score variances determined from outputs of multiple machine learning models. In an embodiment, temporal inconsistency cues are used to select data samples for annotation. In an embodiment, digital map constraints or other map-based data are used to exclude data samples from annotation. In an exemplary application, the annotated data samples are used to train a machine learning model that outputs perception data for an autonomous vehicle application.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Application No. 62 / 796,064, filed on January 23, 2019. Technical Field

[0003] The present specification relates to machine learning, and more particularly to using active learning techniques to automatically select data samples for annotation. Background Art

[0004] Convolutional Neural Networks (CNNs) have been successfully used for perception tasks in autonomous driving applications. The CNN structure consists of layers that are learned through training data. Learning to classify image and video samples requires human annotators to mark each data sample with a label. Having to annotate thousands of data samples is tedious and redundant. Summary of the invention

[0005] An active learning system and method for automatically selecting data samples (e.g., images, point clouds) from a data sample pool for human annotation is disclosed. The system and method select data samples for annotation using bounding box predictions based on bounding box score probability distributions, spatial probability densities determined based on bounding box sizes and positions in data samples, and overall score variances determined based on the outputs of multiple machine learning models. In an embodiment, temporal inconsistency cues are used to select annotated data samples (e.g., consecutive video frames) with temporal noise (e.g., flicker). In an embodiment, a digital map is used to exclude samples that violate map constraints (hard constraints and / or statistical constraints) from annotated data samples. In an exemplary application, annotated data samples are used to train a machine learning model (e.g., CNN) that outputs perceptual data (e.g., labeled objects and scenes) used by an autonomous vehicle application. In an embodiment, the active learning system is implemented in a planning module of an autonomous vehicle.

[0006] Advantages of the disclosed active learning system include automatically selecting a subset of data for human annotation from a large sample pool, thereby reducing the amount of time and cost of manually preparing training data sets for machine learning models for autonomous vehicle applications.

[0007] In an embodiment, a method includes: obtaining, using a computer system, a set of data samples, each data sample including one or more bounding boxes, each bounding box containing a potential object or scene in an environment, each bounding box having a label and a bounding box score indicating a confidence that the label is correct; and selecting, using the computer system, a subset of the data samples for annotation based on bounding box prediction confidences and an overall score variance, wherein the bounding box prediction confidences are determined using a probability distribution of bounding box scores, and the overall score variance is based on differences in overall scores calculated from prediction sets output by multiple machine learning models.

[0008] In an embodiment, the method further comprises: selecting, using the computer system, a subset of the data samples for annotation based on the bounding box prediction confidence, a spatial probability density of the bounding box parameterized by bounding box size and position, and the overall score variance.

[0009] In an embodiment, bounding box prediction further comprises: generating a probability distribution of bounding box scores for each label; and determining a likelihood that a particular bounding box is incorrectly labeled based on the distribution; and selecting the particular bounding box for annotation or excluding the particular bounding box from annotation based on the likelihood.

[0010] In an embodiment, the distribution is approximated by a histogram with bins representing the range of bounding box scores, and each bin is associated with a likelihood.

[0011] In an embodiment, for each interval, the likelihood is calculated according to the ratio of the number of incorrectly labeled bounding boxes assigned to the interval to the number of labeled bounding boxes assigned to the interval.

[0012] In an embodiment, the method further comprises: for each label, sensor and scale, determining the spatial probability density using a Gaussian mixture model (GMM) on a set of bounding boxes for the label, sensor and scale, wherein the GMM is parameterized by bounding box size and position.

[0013] In an embodiment, the spatial probability density of a tag is determined by dividing the spatial density of the tag by the maximum density value among all spatial density values ​​of the tag.

[0014] In an embodiment, the method further includes: processing the data sample through a plurality of different machine learning models to generate predicted labeled bounding boxes; calculating an overall score for each pairwise comparison of the predicted labeled bounding boxes, wherein each predicted labeled bounding box is a ground truth value for comparison with other predicted labeled bounding boxes; and calculating an overall score variance based on the overall score.

[0015] In an embodiment, the plurality of different machine learning models comprises a plurality of different neural networks tuned by training data samples provided by different types of sensors.

[0016] In an embodiment, the different types of sensors include LiDAR, radar, and cameras.

[0017] In an embodiment, the plurality of different neural networks are trained on training data samples in different random orders.

[0018] In one embodiment, the method further comprises: detecting, by the computer system, a temporal inconsistency between consecutive data samples; and selecting the consecutive data samples for annotation based on detecting the temporal inconsistency.

[0019] In an embodiment, the method further comprises: detecting, by the computer system, errors associated with the bounding box using the map constraints; and excluding the bounding box from the annotation based on detecting the errors.

[0020] These and other aspects, features, and implementations are represented as methods, apparatuses, systems, components, program products, means or steps, and other means for performing functions.

[0021] These and other aspects, features and implementations will be apparent from the following description, including the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 An example of an autonomous vehicle with autonomous capabilities is shown.

[0023] Figure 2 An example "cloud" computing environment is illustrated.

[0024] Figure 3 An example computer system is shown.

[0025] Figure 4 An example architecture of an autonomous vehicle is shown.

[0026] Figure 5 Shows examples of inputs and outputs that a perception module can use.

[0027] Figure 6 An example of a LiDAR system is shown.

[0028] Figure 7 The LiDAR system is shown in operation.

[0029] Figure 8 Additional details of the operation of the LiDAR system are shown.

[0030] Fig. 9Block diagram showing the relationship between the inputs and outputs of the planning module.

[0031] Fig.10 Shows the directed graph used in path planning.

[0032] Fig.11 A block diagram showing the inputs and outputs of the control module.

[0033] Fig.12 A block diagram showing the inputs, outputs, and components of the controller.

[0034] Fig.13 Shown is the scene output by the perception module, which includes bounding boxes and corresponding labels and bounding box scores.

[0035] Fig.14 A block diagram of an active learning system for automatically selecting data samples for annotation is shown.

[0036] Fig.15 Shows a stacked histogram of bounding box scores.

[0037] Fig.16 Plot showing the spatial GMM density for a specific tag and a specific sensor.

[0038] Fig.17 A block diagram of the integrated system is shown.

[0039] Fig.18 Flowchart showing the active learning process for automatically selecting samples for annotation. DETAILED DESCRIPTION

[0040] In the following description, for the purpose of explanation, many specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be apparent that the present invention can be implemented without these specific details. In other examples, well-known configurations and devices are shown in block diagram form to avoid unnecessarily obscuring the present invention.

[0041] In the accompanying drawings, for ease of description, a specific arrangement or order of schematic elements (such as those representing devices, modules, instruction blocks, and data elements) is shown. However, it should be understood by those skilled in the art that the specific order or arrangement of the schematic elements in the accompanying drawings is not intended to mean that a specific processing order or sequence, or separation of processing processes is required. In addition, the inclusion of schematic elements in the accompanying drawings is not intended to mean that such elements are required in all embodiments, nor is it intended to mean that the features represented by such elements cannot be included in some embodiments or cannot be combined with other elements in some embodiments.

[0042] In addition, in the accompanying drawings, connecting elements, such as solid or dotted lines or arrows, are used to illustrate the connection, relationship or association between two or more other schematic elements, and there is no such connecting element and is not intended to mean that there can be no connection, relationship or association. In other words, the connection, relationship or association between some elements are not shown in the accompanying drawings, so as not to make the present disclosure vague. In addition, for the convenience of illustration, a single connecting element is used to represent multiple connections, relationships or associations between elements. For example, if the connecting element represents the communication of signal, data or instruction, it will be understood by those skilled in the art that this element represents one or more signal paths (for example, bus) that may be needed to affect the communication.

[0043] 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 embodiments described. However, it will be apparent to one of ordinary skill in the art that the various embodiments described may be implemented without these specific details. In other cases, well-known methods, procedures, components, circuits, and networks are not described in detail in order not to unnecessarily obscure aspects of the embodiments.

[0044] Several features described below can each be used independently of one another, or can be used together with any combination of other features. However, any individual feature may not solve any of the problems discussed above, or may only solve one of the problems discussed above. Some of the problems discussed above may not be fully solved by any one of the features described herein. Although titles are provided, information related to a specific title but not found in the section with the title may also be found elsewhere in this specification. This article describes an embodiment according to the following summary:

[0045] 1. General Overview

[0046] 2. System Overview

[0047] 3. Autonomous Vehicle Architecture

[0048] 4. Autonomous Vehicle Input

[0049] 5. Autonomous vehicle planning

[0050] 6. Autonomous Vehicle Control

[0051] 7. Active Learning System

[0052] General Overview

[0053] We describe techniques for automatically selecting data samples (e.g., images, point clouds) for human annotation, among other things. The techniques select data samples for annotation using bounding box predictions, spatial probability densities determined from bounding box size and position, and overall score variances determined from the outputs of multiple machine learning models. In an embodiment, temporal inconsistency cues are used to select data samples for annotation. In an embodiment, digital map constraints or other map-based data are used to exclude data samples from annotation. In an exemplary application, the annotated data samples are used to train a machine learning model (e.g., CNN), which outputs perceptual data (e.g., labeled objects and scenes) for autonomous vehicle applications. In an embodiment, the active learning system can be implemented by a planning module of an autonomous vehicle.

[0054] System Overview

[0055] Figure 1 An example of an autonomous vehicle 100 having autonomous capabilities is shown.

[0056] As used herein, the term "autonomous capability" refers to a function, feature, or facility that enables a vehicle to operate partially or fully without real-time human intervention, including but not limited to fully autonomous vehicles, highly autonomous vehicles, and conditionally autonomous vehicles.

[0057] As used herein, an autonomous vehicle (AV) is a vehicle with autonomous capabilities.

[0058] As used herein, "vehicle" includes a mode of transport of goods or people. For example, a car, bus, train, airplane, drone, truck, boat, ship, submersible, spacecraft, etc. An unmanned car is an example of a vehicle.

[0059] As used herein, a "track" refers to a path or route that navigates an AV from a first spatiotemporal location to a second spatiotemporal location. In embodiments, the first spatiotemporal location is referred to as an initial location or starting location, and the second spatiotemporal location is referred to as a destination, final location, target, target location, or target location. In some examples, a track consists of one or more road segments (e.g., several sections of a road), and each road segment consists of one or more blocks (e.g., a portion of a lane or intersection). In embodiments, a spatiotemporal location corresponds to a real-world location. For example, a spatiotemporal location is a pickup or drop-off location to get people or cargo on or off the vehicle.

[0060] As used herein, "sensor(s)" includes one or more hardware components for detecting information about the environment surrounding the sensor. Some hardware components may include sensing components (e.g., image sensors, biometric sensors), transmission and / or reception components (e.g., laser or radio frequency wave transmitters and receivers), electronic components (such as analog-to-digital converters), data storage devices (such as RAM and / or non-volatile memory), software or firmware components and data processing components (such as application specific integrated circuits), microprocessors and / or microcontrollers.

[0061] As used herein, a “scene description” is a data structure (e.g., a list) or data stream that includes one or more categorized or labeled objects detected by one or more sensors on an AV vehicle, or provided by a source external to the AV.

[0062] As used herein, a "road" is a physical area that can be traversed by a vehicle and may correspond to a named thoroughfare (e.g., a city street, an interstate highway, etc.) or may correspond to an unnamed thoroughfare (e.g., a driveway within a house or office building, a section of a parking lot, a section of a vacant parking lot, a dirt road in a rural area, etc.). Because some vehicles (e.g., four-wheel drive pickup trucks, off-road vehicles (SUVs), etc.) are able to traverse a variety of physical areas that are not particularly suitable for vehicle travel, a "road" may be any physical area that is not formally defined as a thoroughfare by a municipality or other governmental or administrative agency.

[0063] As used herein, a "lane" is a portion of a road that can be traversed by a vehicle. Lanes are sometimes identified based on lane markings. For example, a lane may correspond to most or all of the space between lane markings, or only some of the space between lane markings (e.g., less than 50%). For example, a road with lane markings that are far apart may accommodate two or more vehicles between the markings, so that one vehicle can pass another vehicle without crossing the lane markings, and thus may be interpreted as having a lane narrower than the space between the lane markings, or having two lanes between the lane markings. Lanes may also be interpreted in the absence of lane markings. For example, lanes may be defined based on physical features of the environment, such as rocks and trees along a pathway in a rural area, or natural obstacles to be avoided in an undeveloped area, for example. Lanes may also be interpreted independently of lane markings or physical features. For example, a lane may be interpreted based on an arbitrary path without obstacles in an area that would otherwise lack features that would be interpreted as lane boundaries. In an example scenario, an AV may interpret a lane through an unobstructed portion of a field or open space. In another example scenario, an AV may interpret a lane through a wide road (e.g., up to two or more lanes wide) without lane markings. In this scenario, the AV can transmit information about the lane to other AVs so that the other AVs can use the same lane information to coordinate path planning between them.

[0064] The term “Over-the-Air (OTA) Client” includes any AV or any electronic device (e.g., computer, controller, IoT device, Electronic Control Unit (ECU)) embedded in, coupled to, or in communication with an AV.

[0065] The term "Over-the-Air (OTA) Update" refers to any update, change, deletion or addition to software, firmware, data or configuration settings, or any combination thereof, delivered to an OTA Client using proprietary and / or standardized wireless communication technologies, including but not limited to: cellular mobile communications (e.g., 2G, 3G, 4G, 5G), radio wireless area networks (e.g., WiFi), and / or satellite Internet.

[0066] The term “edge node” refers to one or more edge devices coupled to a network that provide a portal for communicating with AVs and can communicate with other edge nodes and cloud-based computing platforms for scheduling and delivering OTA updates to OTA clients.

[0067] The term "edge device" refers to a device that implements an edge node and provides a physical wireless access point (AP) to an enterprise or service provider (e.g., VERIZON, AT&T) core network. Examples of edge devices include, but are not limited to: computers, controllers, transmitters, routers, routing switches, integrated access devices (IADs), multiplexers, metropolitan area networks (MANs), and wide area network (WAN) access devices.

[0068] “One or more” includes a function performed by one element, a function performed by multiple elements, such as in a distributed manner, several functions performed by one element, several functions performed by several elements, or any combination of the above.

[0069] It will also be understood that although in some cases, the terms "first," "second," etc. are used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the various 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.

[0070] The terms used in the specification of the various embodiments described herein are only used for the purpose of describing specific embodiments and are not intended to be limiting. As used in the specification of the various embodiments described and the appended claims, the singular forms "a", "an" and "the" are also intended to include plural forms, unless the context clearly indicates otherwise. It will also be understood that "and / or" as used herein refers to and includes any and all possible combinations of one or more related list items. It will also be understood that when the terms "include", "comprises", "have" and / or "have" are used in this specification, the stated features, integers, steps, operations, elements and / or components are specifically stated, but the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof are not excluded.

[0071] As used herein, the term "if" is alternatively understood to mean "when" or "at the time" or "in response to determining that" or "in response to detecting," depending on the context. Similarly, the phrases "if it has been determined" or "if [the stated condition or event] has been detected" are alternatively understood to mean "upon determination" or "in response to determining that" or "upon detecting [the stated condition or event]" or "in response to detecting [the stated condition or event]," depending on the context.

[0072] As used herein, an AV system refers to an AV and the array of hardware, software, stored data, and real-time generated data that supports the operation of the AV. In an embodiment, the AV system is incorporated into the AV. In an embodiment, the AV system is distributed across several locations. For example, some software of the AV system is in a system similar to the following about Figure 2 The described cloud computing environment 200 is implemented in a cloud computing environment.

[0073] In general, this document describes technologies applicable to any vehicle having one or more autonomous capabilities, including fully autonomous vehicles, highly autonomous vehicles, and conditionally autonomous vehicles, such as so-called Level 5, Level 4, and Level 3 vehicles, respectively (see SAE International Standard J3016: Classification and Definitions of Terms Relating to Automated Driving Systems for On-Road Motor Vehicles, which is incorporated by reference in its entirety into this document for more details on vehicle autonomy levels). The technologies described in this document are also applicable to partially autonomous vehicles and driver-assisted vehicles, such as so-called Level 2 and Level 1 vehicles (see SAE International Standard J3016: Classification and Definitions of Terms Relating to Automated Driving Systems for On-Road Motor Vehicles). In embodiments, one or more Level 1, Level 2, Level 3, Level 4, and Level 5 vehicle systems may automatically perform certain vehicle operations (e.g., steering, braking, and using maps) under certain operating conditions based on processing of sensor inputs. The technologies described in this document may benefit any level of vehicle ranging from fully autonomous vehicles to human-operated vehicles.

[0074] Autonomous vehicles have advantages over vehicles that require human drivers. One advantage is safety. For example, in 2016, there were 6 million car accidents in the United States, 2.4 million injuries, 40,000 deaths, 13 million vehicle collisions, and an estimated societal cost of $910+ billion. From 1965 to 2015, traffic fatalities in the United States have decreased from about 6 to about 1 per 100 million miles driven, in part due to the additional safety measures deployed in vehicles. For example, an extra half-second of warning before a collision is expected is thought to mitigate 60% of front-to-rear collisions. However, passive safety features (e.g., seat belts, airbags) may have reached their limits in improving these numbers. Therefore, active safety measures such as automated control of vehicles may be the next step in improving these statistics. Because human drivers are considered responsible for severe pre-crash accidents in 95% of crashes, autonomous driving systems may achieve better safety outcomes, such as identifying and avoiding critical situations more reliably than humans; making better decisions, obeying traffic rules, and predicting future events better than humans; and controlling vehicles more reliably than humans.

[0075] refer to Figure 1 , the AV system 120 causes the AV 100 to operate along a trajectory 198, through an environment 190 to a destination 199 (sometimes referred to as a final location), while avoiding objects (e.g., natural obstacles 191, vehicles 193, pedestrians 192, cyclists, and other obstacles) and complying with road rules (e.g., operating rules or driving preferences).

[0076] In an embodiment, the AV system 120 includes a device 101 equipped to receive and operate operating commands from a computer processor 146. In an embodiment, the computing processor 146 is associated with the following reference Figure 3 The processor 304 is similarly described. Examples of devices 101 include steering controls 102, brakes 103, gears, an accelerator pedal or other acceleration control mechanism, windshield wipers, side door locks, window controls, and turn indicators.

[0077] In an embodiment, the AV system 120 includes sensors 121 for measuring or inferring attributes of the state or condition of the AV 100, such as the position, linear and angular velocity and acceleration, and heading (e.g., the direction of the front end of the AV 100) of the AV. Examples of sensors 121 are GPS, an inertial measurement unit (IMU) that measures both linear acceleration and angular rate of the vehicle, wheel rate sensors for measuring or estimating wheel slip, wheel brake pressure or brake torque sensors, engine torque or wheel torque sensors, and steering angle and angular rate sensors.

[0078] In an embodiment, the sensors 121 also include sensors for sensing or measuring properties of the environment of the AV, such as monocular or stereo cameras 122 in the visible, infrared, or thermal (or both) spectrum, LiDAR 123, radar, ultrasonic sensors, time-of-flight (TOF) depth sensors, velocity sensors, temperature sensors, humidity sensors, and precipitation sensors.

[0079] In an embodiment, the AV system 120 includes a data storage unit 142 and a memory 144 for storing machine instructions associated with a computer processor 146 or data collected by the sensor 121. In an embodiment, the data storage unit 142 is associated with the following Figure 3 190 . In an embodiment, the memory 144 is similar to the main memory 306 described below. In an embodiment, the data storage unit 142 and the memory 144 store historical, real-time, and / or predictive information about the environment 190. In an embodiment, the stored information includes maps, driving performance, traffic congestion updates, or weather conditions. In an embodiment, data related to the environment 190 is transmitted from the remote database 134 to the AV 100 via a communication channel.

[0080] In an embodiment, the AV system 120 includes a communication device 140 for transmitting measured or inferred properties of the state and condition of other vehicles (such as position, linear and angular velocity, linear and angular acceleration, and linear and angular heading) to the AV 100. These devices include vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication devices and devices for wireless communication through point-to-point or ad hoc networks or both. In an embodiment, the communication device 140 communicates across the electromagnetic spectrum (including radio and optical communications) or other media (e.g., air and acoustic media). The combination of vehicle-to-vehicle (V2V), vehicle-to-infrastructure (V2I) communications (and in some embodiments, one or more other types of communications) is sometimes referred to as vehicle-to-everything (V2X) communications. V2X communications typically comply with one or more communication standards for communication with and between autonomous vehicles.

[0081] In an embodiment, the communication device 140 includes a communication interface. For example, a wired, wireless, WiMAX, WiFi, Bluetooth, satellite, cellular, optical, near field, infrared, or radio interface. The communication interface transmits data from the remote database 134 to the AV system 120. In an embodiment, the remote database 134 is embedded in a Figure 2 The communication interface 140 transmits data collected from the sensor 121 or other data related to the operation of the AV 100 to the remote database 134. In an embodiment, the communication interface 140 transmits information related to teleoperation to the AV 100. In some embodiments, the AV 100 communicates with other remote (e.g., "cloud") servers 136.

[0082] In an embodiment, the remote database 134 also stores and transmits digital data (e.g., stores data such as roads and street locations). Such data is stored in the memory 144 on the AV 100 or transmitted from the remote database 134 to the AV 100 via a communication channel.

[0083] In an embodiment, the remote database 134 stores and transmits historical information (e.g., speed and acceleration profiles) related to driving attributes of vehicles that have previously traveled along the trajectory 198 at similar times of day. In one implementation, such data may be stored in the memory 144 on the AV 100 or transmitted from the remote database 134 to the AV 100 via a communication channel.

[0084] The computing device 146 located on the AV 100 algorithmically generates control actions based on both real-time sensor data and a priori information, allowing the AV system 120 to perform its autonomous driving capabilities.

[0085] In an embodiment, the AV system 120 includes a computer peripheral device 132 coupled to the computing device 146 for providing information and reminders to a user of the AV 100 (e.g., an occupant or a remote user) and receiving input from the user. In an embodiment, the peripheral device 132 is similar to the one described below with reference to Figure 3 Display 312, input device 314 and cursor control 316 are discussed. The coupling may be wireless or wired. Any two or more of the interface devices may be integrated into a single device.

[0086] Figure 2 An example "cloud" computing environment is illustrated. Cloud computing is a service delivery model for enabling convenient, on-demand access to a shared pool of configurable computing resources (e.g., networks, network bandwidth, servers, processing, memory, storage, applications, virtual machines, and services) over a network. In a typical cloud computing system, one or more large cloud data centers house the machines used to deliver the services provided by the cloud. Now refer to Figure 2 , cloud computing environment 200 includes cloud data centers 204a, 204b, and 204c interconnected by cloud 202. Data centers 204a, 204b, and 204c provide cloud computing services for computer systems 206a, 206b, 206c, 206d, 206e, and 206f connected to cloud 202.

[0087] The cloud computing environment 200 includes one or more cloud data centers. Figure 2 The cloud data center 204a shown in FIG. 2 refers to a cloud (eg, Figure 2 The physical arrangement of servers in a cloud 202 or a specific portion of a cloud (as shown in FIG. 202 or a specific portion of a cloud). For example, servers are physically arranged into rooms, groups, rows, and racks in a cloud data center. A cloud data center has one or more zones, which include one or more server rooms. Each room has one or more rows of servers, and each row includes one or more racks. Each rack includes one or more individual server nodes. In some implementations, servers in zones, rooms, racks, and / or rows are arranged into groups based on the physical infrastructure requirements of the data center facility, including power, energy, heat, heat sources, and / or other requirements. In an embodiment, server nodes are similar to Figure 3 The data center 204a has many computing systems distributed across multiple racks.

[0088] Cloud 202 includes cloud data centers 204a, 204b, and 204c and networks and network resources (e.g., network devices, nodes, routers, switches, and network cables) used to connect cloud data centers 204a, 204b, and 204c and help facilitate access to cloud computing services by computing systems 206a-f. In an embodiment, the network represents any combination of one or more local networks, wide area networks, or internetworks coupled by wired or wireless links deployed using terrestrial or satellite connections. Data exchanged through the network is transmitted using a variety of network layer protocols (such as Internet Protocol (IP), Multi-Protocol Label Switching (MPLS), Asynchronous Transfer Mode (ATM), Frame Relay, etc.). In addition, in embodiments where the network represents a combination of multiple subnetworks, different network layer protocols are used on each underlying subnetwork. In some embodiments, the network represents one or more interconnected internetworks (such as the public Internet, etc.).

[0089] Computing systems 206a-f or cloud computing service consumers are connected to the cloud 202 through network links and network adapters. In an embodiment, computing systems 206a-f are implemented as various computing devices, such as servers, desktops, laptops, tablets, smart phones, Internet of Things (IoT) devices, autonomous vehicles (including cars, drones, shuttles, trains, buses, etc.) and consumer electronics. In an embodiment, computing systems 206a-f are implemented in other systems or as part of other systems.

[0090] Figure 3 Computer system 300 is illustrated. In implementation, computer system 300 is a special purpose computing device. The special purpose computing device is hardwired to perform these techniques, or includes a digital electronic device such as one or more application specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs) that are permanently programmed to perform the above-mentioned techniques, or may include one or more general purpose hardware processors that are programmed to perform these techniques according to program instructions in firmware, memory, other memory, or a combination. Such a special purpose computing device can also combine customized hardwired logic, ASICs or FPGAs with customized programming to complete these techniques. In various embodiments, the special purpose computing device is a desktop computer system, a portable computer system, a handheld device, a network device, or any other device that contains hardwired and / or program logic to implement these techniques.

[0091] In an embodiment, computer system 300 includes a bus 302 or other communication mechanism for communicating information, and a hardware processor 304 coupled to bus 302 to process information. Hardware processor 304 is, for example, a general-purpose microprocessor. Computer system 300 also includes a main memory 306, such as a random access memory (RAM) or other dynamic storage device, coupled to bus 302 to store information and instructions, which are executed by processor 304. In one implementation, main memory 306 is used to store temporary variables or other intermediate information during the execution of instructions to be executed by processor 304. When these instructions are stored in a non-transitory storage medium accessible to processor 304, computer system 300 becomes a special-purpose machine that is customized to perform the operations specified in the instructions.

[0092] In an embodiment, computer system 300 also includes a read only memory (ROM) 308 or other static storage device coupled to bus 302 for storing static information and instructions for processor 304. A storage device 310, such as a magnetic disk, optical disk, solid state drive, or three-dimensional cross point memory, is provided and coupled to bus 302 to store information and instructions.

[0093] In an embodiment, the computer system 300 is coupled to a display 312 such as a cathode ray tube (CRT), a liquid crystal display (LCD), a plasma display, a light emitting diode (LED) display, or an organic light emitting diode (OLED) display for displaying information to a computer user via bus 302. An input device 314 including alphanumeric and other keys is coupled to bus 302 for communicating information and command selections to processor 304. Another type of user input device is a cursor controller 316, such as a mouse, a trackball, a touch display, or cursor direction keys, for communicating direction information and command selections to processor 304 and for controlling movement of a cursor on display 312. Such input devices typically have two degrees of freedom in two axes, a first axis (e.g., an x-axis) and a second axis (e.g., a y-axis) that allow the device to specify a position on a plane.

[0094] According to one embodiment, the techniques herein are performed by computer system 300 in response to processor 304 executing one or more sequences of one or more instructions contained in main memory 306. These instructions are read into main memory 306 from another storage medium, such as storage device 310. Execution of the sequences of instructions contained in main memory 306 causes processor 304 to perform the process steps described herein. In alternative embodiments, hard-wired circuitry is used in place of or in combination with software instructions.

[0095] As used herein, the term "storage medium" refers to any non-temporary medium that stores data and / or instructions that cause a machine to operate in a particular manner. Such storage media include non-volatile media and / or volatile media. Non-volatile media include, for example, optical disks, magnetic disks, solid-state drives, or three-dimensional cross-point memories such as storage device 310. Volatile media include dynamic memories such as main memory 306. Common forms of storage media include, for example, floppy disks, floppy disks, hard disks, solid-state drives, tapes or any other magnetic data storage media, CD-ROMs, any other optical data storage media, any physical media with a hole pattern, RAM, PROM and EPROM, FLASH-EPROM, NV-RAM, or any other memory chip or storage box.

[0096] Storage media are distinct from transmission media, but can be used in conjunction with transmission media. Transmission media participate in the transmission of information between storage media. For example, transmission media include coaxial cables, copper wires, and optical fibers, including wires that provide bus 302. Transmission media can also take the form of sound waves or light waves, such as those generated during radio wave and infrared data communications.

[0097] In an embodiment, various forms of media are involved in carrying one or more sequences of one or more instructions to the processor 304 for execution. For example, the instructions are initially executed on a disk or solid-state drive of a remote computer. The remote computer loads the instructions into its dynamic memory and sends the instructions over a telephone line using a modem. The local modem of the computer system 300 receives the data on the telephone line and uses an infrared transmitter to convert the data to an infrared signal. An infrared detector receives the data carried in the infrared signal, and appropriate circuitry places the data on the bus 302. The bus 302 carries the data to the main memory 306, from which the processor 304 retrieves and executes the instructions. The instructions received by the main memory 306 may optionally be stored on the storage device 310 before or after execution by the processor 304.

[0098] Computer system 300 also includes a communication interface 318 coupled to bus 302. Communication interface 318 provides a two-way data communication coupled to a network link 320 connected to a local network 322. For example, communication interface 318 is an integrated services digital network (ISDN) card, a cable modem, a satellite modem, or a modem for providing a data communication connection with a corresponding type of telephone line. As another example, communication interface 318 is a local area network (LAN) card for providing a data communication connection with a compatible LAN. In some implementations, a wireless link is also implemented. In any such implementation, communication interface 318 sends and receives an electrical, electromagnetic or optical signal carrying a digital data stream representing various types of information.

[0099] The network link 320 typically provides data communication to other data devices through one or more networks. For example, the network link 320 provides a connection to a host computer 324 or to a cloud data center or device operated by an Internet Service Provider (ISP) 326 through a local network 322. The ISP 326, in turn, provides data communication services through a worldwide packet data communication network now commonly referred to as the "Internet" 328. Both the local network 322 and the Internet 328 use electrical, electromagnetic, or optical signals that carry digital data streams. The signals through the various networks and the signals on the network link 320 and through the communication interface 318 are example forms of transmission media, where these signals carry digital data to and from the computer system 300. In an embodiment, the network 320 includes the cloud 202 described above or a portion of the cloud 202.

[0100] Computer system 300 sends messages and receives data, including program code, through network(s), network link 320, and communication interface 318. In an embodiment, computer system 300 receives code for processing. The received code is executed by processor 304 as it is received, and / or stored in storage device 310, or other non-volatile storage for later execution.

[0101] Autonomous Vehicle Architecture

[0102] Figure 4 An example of a method for autonomous vehicles (e.g., Figure 1 100). The architecture 400 includes a sensing module 402 (sometimes referred to as sensing circuitry), a planning module 404 (sometimes referred to as planning circuitry), a control module 406 (sometimes referred to as control circuitry), a positioning module 408 (sometimes referred to as positioning circuitry), and a database module 410 (sometimes referred to as database circuitry). Each module plays a role in the operation of the AV 100. Collectively, the modules 402, 404, 406, 408, and 410 may be Figure 1 404, 406, 408, and 410 are a portion of the AV system 120 shown. In some embodiments, any of the modules 402, 404, 406, 408, and 410 is a combination of computer software (e.g., executable code stored on a computer-readable medium) and computer hardware (e.g., one or more microprocessors, microcontrollers, application specific integrated circuits [ASICs], hardware memory devices, other types of integrated circuits, other types of computer hardware, or any or all combinations of these hardware). Individual modules in the modules 402, 404, 406, 408, and 410 are sometimes referred to as processing circuits (e.g., computer hardware, computer software, or a combination of both). The combination of any or all of the modules 402, 404, 406, 408, and 410 is also an example of a processing circuit.

[0103] In use, planning module 404 receives data representing destination 412 and determines data representing a trajectory 414 (sometimes referred to as a route) that AV 100 may travel in order to reach (e.g., arrive at) destination 412. In order for planning module 404 to determine data representing trajectory 414, planning module 404 receives data from perception module 402, positioning module 408, and database module 410.

[0104] The perception module 402 uses, for example, Figure 1 One or more sensors 121 shown are used to identify nearby physical objects. Objects are classified (e.g., grouped into types such as pedestrians, bicycles, cars, traffic signs, etc.), and a scene description including the classified objects 416 is provided to the planning module 404. In an embodiment, the perception module 402 includes an object detector that detects objects and labels the objects. An example object detector is a convolutional neural network (CNN). In an embodiment, the object detector can output an image or a point cloud that includes a bounding box around the detected object, a label for the object, and a bounding box score indicating the confidence that the detected object is correctly labeled. In an embodiment, the bounding box score can be in the range [0.0 1.0], where 0.0 indicates an incorrectly labeled object, 1.0 indicates a correctly labeled object, and values ​​between the two indicate the confidence that the object is correctly labeled.

[0105] The planning module 404 also receives data representing the AV position 418 from the positioning module 408. The positioning module 408 determines the AV position by using data from the sensor 121 and data from the database module 410 (e.g., geographic data) to calculate the position. For example, the positioning module 408 uses data from a GNSS (Global Navigation Satellite System) sensor and geographic data to calculate the longitude and latitude of the AV. In an embodiment, the data used by the positioning module 408 includes a high-precision map with lane geometry attributes, a map describing the road network connection attributes, a map describing the physical attributes of the lane (such as traffic speed, traffic volume, the number of vehicles and bicycle lanes, lane width, lane traffic direction, or lane marking type and location, or a combination thereof), and a map describing the spatial location of road features (such as intersections, traffic signs or various types of other driving signals, etc.). In an embodiment, the high-precision map is constructed by adding data to the low-precision map through automatic or manual annotations.

[0106] The control module 406 receives data representing the trajectory 414 and data representing the AV position 418, and operates control functions 420a-420c (e.g., steering, throttle, brakes, ignition) of the AV in a manner that will cause the AV 100 to travel the trajectory 414 to reach the destination 412. For example, if the trajectory 414 includes a left turn, the control module 406 will operate the control functions 420a-420c in the following manner: the steering angle of the steering function will cause the AV 100 to turn left, and the throttle and brakes will cause the AV 100 to pause and wait for a passing pedestrian or vehicle before making the turn.

[0107] Autonomous Vehicle Input

[0108] Figure 5 The perception module 402 ( Figure 4 ) used by the inputs 502a-502d (e.g., Figure 1 1) and outputs 504a-504d (e.g., sensor data). One input 502a is a LiDAR (Light Detection and Ranging) system (e.g., Figure 1 123). LiDAR is a technology that uses light (e.g., a beam of light such as infrared light) to obtain data about physical objects in its line of sight. The LiDAR system produces LiDAR data as output 504a. For example, LiDAR data is a collection of 3D or 2D points (also called a point cloud) used to construct a representation of the environment 190.

[0109] Another input 502b is a RADAR system. Radar is a technology that uses radio waves to obtain data about nearby physical objects. Radar can obtain data about objects that are not within the line of sight of the LiDAR system. Radar system 502b produces radar data as output 504b. For example, radar data is one or more radio frequency electromagnetic signals used to construct a representation of environment 190.

[0110] Another input 502c is a camera system. The camera system uses one or more cameras (e.g., a digital camera using a light sensor such as a charge coupled device [CCD]) to obtain information about nearby physical objects. The camera system generates camera data as output 504c. The camera data is typically in the form of image data (e.g., data in an image data format such as RAW, JPEG, PNG, etc.). In some examples, the camera system has multiple independent cameras, such as for the purpose of stereoscopic imaging (stereoscopic vision), which enables the camera system to perceive depth. Although the objects perceived by the camera system are described here as "nearby", this is relative to the AV. In use, the camera system can be configured to "see" distant objects (e.g., as far as 1 km or more in front of the AV). Therefore, the camera system can have features such as sensors and lenses that are optimized for perceiving distant objects.

[0111] Another input 502d is a traffic light detection (TLD) system. The TLD system uses one or more cameras to obtain information about traffic lights, street signs, and other physical objects that provide visual navigation information. The TLD system generates TLD data as an output 504d. The TLD data often takes the form of image data (e.g., data in image data formats such as RAW, JPEG, PNG, etc.). The TLD system differs from a system that includes a camera in that the TLD system uses a camera with a wide field of view (e.g., using a wide-angle lens or a fisheye lens) to obtain information about as many physical objects that provide visual navigation information as possible, so that the AV 100 can access all relevant navigation information provided by these objects. For example, the viewing angle of the TLD system can be about 120 degrees or greater.

[0112] In some embodiments, sensor fusion techniques are used to combine the outputs 504a-504d. Thus, the individual outputs 504a-504d are provided to other systems of the AV 100 (e.g., to a Figure 4 The combined output may be provided to other systems in the form of a single combined output or multiple combined outputs of the same type (e.g., using the same combining technique or combining the same outputs, or both) or a single combined output or multiple combined outputs of different types (e.g., using different individual combining techniques or combining different individual outputs, or both). In some embodiments, an early fusion technique is used. An early fusion technique is characterized in that the outputs are combined before one or more data processing steps are applied to the combined output. In some embodiments, a late fusion technique is used. A late fusion technique is characterized in that the outputs are combined after one or more data processing steps are applied to the individual outputs.

[0113] Figure 6 An example of a LiDAR system 602 is shown (eg, Figure 5 502a shown). The LiDAR system 602 emits light 604a-604c from a light emitter 606 (e.g., a laser emitter). The light emitted by the LiDAR system is typically not in the visible spectrum; for example, infrared light is often used. Some of the emitted light 604b encounters a physical object 608 (e.g., a vehicle) and reflects back to the LiDAR system 602. (The light emitted from the LiDAR system typically does not penetrate a physical object, such as a physical object that is solid in form.) The LiDAR system 602 also has one or more light detectors 610 for detecting the reflected light. In an embodiment, one or more data processing systems associated with the LiDAR system generate an image 612 representing a field of view 614 of the LiDAR system. The image 612 includes information representing a boundary 616 of the physical object 608. In this way, the image 612 is used to determine the boundaries 616 of one or more physical objects near the AV.

[0114] Figure 7 6. The LiDAR system 602 is shown in operation. In the scenario shown in the figure, the AV 100 receives both the camera system output 504c in the form of an image 702 and the LiDAR system output 504a in the form of LiDAR data points 704. In use, the data processing system of the AV 100 compares the image 702 to the data points 704. In particular, physical objects 706 identified in the image 702 are also identified in the data points 704. In this way, the AV 100 perceives the boundaries of the physical objects based on the contours and density of the data points 704.

[0115] Figure 8 6 shows additional details of the operation of the LiDAR system 602. As described above, the AV 100 detects the boundaries of the physical object based on the characteristics of the data points detected by the LiDAR system 602. Figure 8 As shown, a flat object such as the ground 802 will reflect light 804a-804d emitted from the LiDAR system 602 in a consistent manner. In other words, because the LiDAR system 602 emits light using consistent intervals, the ground 802 will reflect light back to the LiDAR system 602 at the same consistent intervals. As the AV 100 drives over the ground 802, the LiDAR system 602 will continue to detect light reflected by the next valid ground point 806 if nothing is blocking the road. However, if an object 808 is blocking the road, the light 804e-804f emitted by the LiDAR system 602 will be reflected from points 810a-810b in a manner that is inconsistent with the expected consistent manner. Based on this information, the AV 100 can determine that the object 808 is present.

[0116] Path Planning

[0117] Fig. 9 Shown (for example, Figure 4 900 of the relationship between the input and output of the planning module 404 (shown in FIG. 1 ). In general, the output of the planning module 404 is a route 902 from a starting point 904 (e.g., a source location or initial location) to an end point 906 (e.g., a destination or final location). The route 902 is typically defined by one or more road segments. For example, a road segment refers to a distance to be traveled on at least a portion of a street, road, highway, lane, or other physical area suitable for automobile travel. In some examples, for example, if the AV 100 is an off-road capable vehicle such as a four-wheel drive (4WD) or all-wheel drive (AWD) car, SUV, or pickup truck, the route 902 includes "off-road" road segments such as unpaved paths or open fields.

[0118] In addition to the route 902, the planning module also outputs lane-level route planning data 908. The lane-level route planning data 908 is used to drive through the road segments of the route 902 at a specific time based on the conditions of the road segments. For example, if the route 902 includes a multi-lane highway, the lane-level route planning data 908 includes trajectory planning data 910, wherein the AV 100 can use the trajectory planning data 910 to select a lane from the multiple lanes based on, for example, whether an exit is approaching, whether there are other vehicles in one or more of the multiple lanes, or other factors that change over the course of a few minutes or less. Similarly, in some implementations, the lane-level route planning data 908 includes a speed constraint 912 that is specific to a road segment of the route 902. For example, if the road segment includes pedestrians or unexpected traffic, the speed constraint 912 can limit the AV 100 to a travel speed that is slower than the expected speed, such as a speed based on the speed limit data of the road segment.

[0119] In an embodiment, inputs to the planning module 404 include (e.g., Figure 4 ) database data 914, current location data 916 (e.g., Figure 4 AV position 418 shown), (e.g., for Figure 4 Destination data 918 and object data 920 (e.g., such as the destination 412 shown) Figure 44 (e.g., a classified object 416 perceived by the perception module 402 shown in FIG. 4 ). In some embodiments, the database data 914 includes rules used in planning. The rules are specified using a formal language (e.g., using Boolean logic). In any given situation encountered by the AV 100, at least some of these rules will apply to the situation. A rule applies to a given situation if it has a condition that is satisfied based on information available to the AV 100 (e.g., information about the surrounding environment). Rules can have priorities. For example, a rule of "if the road is a highway, move to the leftmost lane" can have a lower priority than "if the exit is within one mile, move to the rightmost lane."

[0120] Fig.10 In path planning (eg, by planning module 404 ( Figure 4 )) uses a directed graph 1000. In general, if Fig.10 A directed graph 1000 such as the one shown is used to determine a path between any origin 1002 and destination 1004. In the real world, the distance separating the origin 1002 and destination 1004 may be relatively large (e.g., in two different metropolitan areas), or may be relatively small (e.g., two intersections adjacent to a city block or two lanes of a multi-lane road).

[0121] In an embodiment, directed graph 1000 has nodes 1006a-1006d representing different locations that AV 100 may occupy between starting point 1002 and end point 1004. In some examples, for example, when starting point 1002 and end point 1004 represent different metropolitan areas, nodes 1006a-1006d represent sections of a road. In some examples, for example, when starting point 1002 and end point 1004 represent different locations on the same road, nodes 1006a-1006d represent different locations on the road. In this way, directed graph 1000 includes information at different granularity levels. In an embodiment, a directed graph with high granularity is also a subgraph of another directed graph with a larger scale. For example, most of the information for a directed graph where the start point 1002 and the end point 1004 are far apart (e.g., many miles apart) is at a low granularity and the directed graph is based on stored data, but the directed graph also includes some high granularity information for a portion of the directed graph that represents a physical location in the field of view of the AV 100.

[0122] Nodes 1006a-1006d are different from objects 1008a-1008b that cannot overlap with nodes. In an embodiment, at low granularity, objects 1008a-1008b represent areas that cars cannot pass through, such as areas without streets or roads. At high granularity, objects 1008a-1008b represent physical objects in the field of view of AV 100, such as other cars, pedestrians, or other entities with which AV 100 cannot share physical space. In an embodiment, some or all of objects 1008a-1008b are static objects (e.g., objects that do not change position, such as street lights or telephone poles, etc.) or dynamic objects (e.g., objects that can change position, such as pedestrians or other cars, etc.).

[0123] Nodes 1006a-1006d are connected by edges 1010a-1010c. If two nodes 1006a-1006b are connected by edge 1010a, the AV 100 can travel between one node 1006a and another node 1006b, for example, without having to travel to an intermediate node before reaching the other node 1006b. (When referring to the AV 100 traveling between nodes, it means that the AV 100 travels between two physical locations represented by the corresponding nodes.) Edges 1010a-1010c are generally bidirectional, in the sense that the AV 100 travels from a first node to a second node, or from a second node to a first node. In an embodiment, edges 1010a-1010c are unidirectional, in the sense that the AV 100 can travel from a first node to a second node, but the AV 100 cannot travel from a second node to the first node. Edges 1010a-1010c are unidirectional where edges 1010a-1010c represent, for example, a one-way street, a single lane of a street, road, or highway, or other features that can only be traversed in one direction due to legal or physical constraints.

[0124] In an embodiment, the planning module 404 uses the directed graph 1000 to identify a path 1012 consisting of nodes and edges between the start point 1002 and the end point 1004 .

[0125] Edges 1010a-1010c have associated costs 1014a-1014b. Costs 1014a-1014b are values ​​representing resources that would be spent if the AV 100 selected that edge. A typical resource is time. For example, if the physical distance represented by one edge 1010a is twice the physical distance represented by another edge 1010b, the associated cost 1014a of the first edge 1010a may be twice the associated cost 1014b of the second edge 1010b. Other factors that affect time include expected traffic, the number of intersections, speed limits, etc. Another typical resource is fuel economy. Two edges 1010a-1010b may represent the same physical distance, but one edge 1010a requires more fuel than the other edge 1010b, for example, due to road conditions, expected weather, etc.

[0126] When the planning module 404 identifies a path 1012 between the start point 1002 and the end point 1004 , the planning module 404 typically selects a path that is optimized for cost, eg, a path that has a minimum total cost when the individual costs of the edges are added together.

[0127] Autonomous Vehicle Control

[0128] Fig.11 Shown (for example, Figure 4 1100 of the inputs and outputs of the control module 406 (shown in FIG. 110). The control module operates according to a controller 1102, which includes, for example: one or more processors similar to the processor 304 (e.g., one or more computer processors such as a microprocessor or a microcontroller or both); short-term and / or long-term data storage devices similar to the main memory 306, ROM 308, and storage devices 310 (e.g., memory, random access memory, or flash memory or both); and instructions stored in the memory that, when executed (e.g., by the one or more processors), perform the operations of the controller 1102.

[0129] In an embodiment, the controller 1102 receives data representing a desired output 1104. The desired output 1104 typically includes speed, such as velocity and heading. The desired output 1104 may be based on, for example, Figure 4104 . Based on the desired output 1104, the controller 1102 generates data that can be used as a throttle input 1106 and a steering input 1108. The throttle input 1106 represents the amount by which the throttle of the AV 100 (e.g., an acceleration control) is engaged to achieve the desired output 1104, such as by engaging a steering pedal or engaging another throttle control. In some examples, the throttle input 1106 also includes data that can be used to engage the brakes of the AV 100 (e.g., a deceleration control). The steering input 1108 represents a steering angle, such as the angle at which the steering control of the AV (e.g., a steering wheel, a steering angle actuator, or other function for controlling the steering angle) should be positioned to achieve the desired output 1104.

[0130] In an embodiment, the controller 1102 receives feedback used in adjusting the inputs provided to the throttle and steering. For example, if the AV 100 encounters a disturbance 1110, such as a hill, the measured velocity 1112 of the AV 100 drops below the desired output rate. In an embodiment, any measured output 1114 is provided to the controller 1102 so that the required adjustments can be made, for example, based on the difference 1113 between the measured velocity and the desired output. The measured output 1114 includes measured position 1116, measured velocity 1118 (including velocity and heading), measured acceleration 1120, and other outputs that the sensors of the AV 100 can measure.

[0131] In an embodiment, information about the disturbance 1110 is detected in advance, for example, by a sensor such as a camera or LiDAR sensor, and provided to the predictive feedback module 1122. The predictive feedback module 1122 then provides information to the controller 1102 that the controller 1102 can use to adjust accordingly. For example, if the sensors of the AV 100 detect ("see") a hill, the controller 1102 can use this information to prepare to engage the throttle at an appropriate time to avoid significant deceleration.

[0132] Fig.12 A block diagram 1200 is shown of the inputs, outputs, and components of the controller 1102. The controller 1102 has a rate analyzer 1206 that affects the operation of a throttle / brake controller 1204. For example, the rate analyzer 1202 instructs the throttle / brake controller 1204 to use a throttle / brake 1206 to accelerate or decelerate based on, for example, feedback received by the controller 1102 and processed by the rate analyzer 1202.

[0133] Controller 1102 also has a lateral tracking controller 1208 that affects the operation of steering wheel controller 1210. For example, lateral tracking controller 1208 instructs steering wheel controller 1210 to adjust the position of steering angle actuator 1212 based on feedback received by controller 1102 and processed by lateral tracking controller 1208, for example.

[0134] The controller 1102 receives several inputs used to determine how to control the throttle / brake 1206 and the steering angle actuator 1212. The planning module 404 provides information used by the controller 1102, for example, to select a heading for the AV 100 to begin operation and to determine which road segment to cross when the AV 100 reaches an intersection. The positioning module 408 provides information describing the current location of the AV 100 to the controller 1102, for example, so that the controller 1102 can determine whether the AV 100 is in the expected location based on the way the throttle / brake 1206 and the steering angle actuator 1212 are being controlled. In an embodiment, the controller 1102 receives information from other inputs 1214, such as information received from a database, a computer network, etc.

[0135] Active learning system

[0136] Fig.13 An example data sample 1300 output by the perception module 402 is shown, which includes 10 bounding boxes 1301a-1301k with corresponding bounding box labels and scores. In the example shown, the data sample is an image captured by a camera. The bounding box label identifies the object, and the bounding box score indicates the confidence level that its corresponding bounding box includes the correctly labeled object. In any given data sample, there may be bounding boxes that include objects, bounding boxes that do not include objects, and bounding boxes that may include objects. In data sample 1300, the bounding box score ranges from 0.00 to 1.00, where 0.00 means that the object is not detected, 1.00 means that the object is detected and correctly labeled, and the numbers between the two represent the confidence that the object is detected and correctly labeled. For example, bounding box 1301c contains a car and has a bounding box score of 1.00, which indicates that the car is detected and correctly labeled as "car". Likewise, bounding box 1301d includes a pedestrian and has a bounding box score of 0.96, indicating a high confidence that the pedestrian is correctly labeled as a "person". Bounding box 1301k does not contain an object and therefore has a bounding box score of 0.00. Note that in practice, each bounding box has a bounding box score for each category. However, after the various filtering steps, only bounding boxes with high scores (above a specified threshold) are included in the output of the perception module.

[0137] Thousands of data samples with bounding boxes and bounding box scores can be collected from the perception module in the AV and stored in a database. In order to reduce the cost and time of manual labeling, as shown below Fig.14 As described, an active learning system is used to automatically ask a human annotator which unlabeled bounding box is to be labeled or which incorrectly labeled bounding box is to be relabeled. In an embodiment, the active learning system is implemented in the perception module 402 of the AV 100, and the selected data samples are provided to the human annotator by the computer system 300 of the AV 100 via wireless transmission or other transmission mechanism (e.g., thumb drive). In other embodiments, the AV 100 provides the collected data samples to the human annotator, and an active learning application installed on a computer system remote from the AV 100 is used to automatically select data samples for annotation.

[0138] Since there is too much sensor data to store in actual systems, the active learning system disclosed herein helps reduce the amount of data samples stored in the database by identifying erroneous annotations in the database, and finds a balance between the quantity and quality of data samples in the database to optimize system performance and cost. In addition, the impact of different kinds of annotation errors (e.g., missing objects, inaccurate locations) on performance can be analyzed, and the same neural network can be trained with different amounts of data.

[0139] Fig.14 A block diagram of an active learning system 1400 for automatically selecting a subset of data samples 1401 for annotation is shown. In the example shown, the active learning system 1400 includes data samples 1401, a bounding box prediction module 1402, a spatial probability density module 1403, an overall score variance module 1404, selected data samples 1405, and annotators 1406.

[0140] There are scenarios where unlabeled or incorrectly labeled data is plentiful but manual labeling by humans is expensive. In such scenarios, the active learning system 1400 automatically asks annotators 1406 for labels. Data samples 1401 (e.g., images, point clouds) include bounding boxes with known labels and bounding boxes with unknown labels. The labeled bounding boxes may be correctly labeled or incorrectly labeled. The active learning system 1400 automatically selects a subset of bounding boxes with unknown and / or incorrect labels to be manually labeled (annotated) by annotators 1406.

[0141] More specifically, in a first step, the bounding box prediction module 1402 retrieves a set of labeled bounding boxes and their corresponding bounding box scores from the input data sample 1401. The bounding box prediction module 1402 then generates a probability distribution of bounding box scores over the set of labeled bounding boxes. In an embodiment, the probability distribution is approximated by a stacked histogram, such as Fig.15 shown.

[0142] Fig.15 A stacked histogram of bounding box scores is shown. For each label, a stacked histogram is generated based on the number of correctly labeled and incorrectly labeled bounding boxes in the labeled bounding box set. In this example, the stacked histogram is for the "pedestrian" label. The x-axis includes intervals of the histogram representing the range (0.0 to 1.0) of the bounding box score. The y-axis is the number of bounding boxes in each interval. Each interval has the number of incorrectly labeled bounding boxes 1501 and the number of correctly labeled bounding boxes 1502. These numbers depend on the optimal confidence threshold 1503 and the ground truth label. The optimal confidence threshold 1503 weights the relationship between false positives (too many detections) and false negatives (too many missing detections). For example, a bounding box with a score that falls to the left of the optimal confidence threshold 1503 is considered to have an incorrect label, while a bounding box with a score that falls to the right of the optimal confidence threshold 1503 is considered to have a correct label.

[0143] The likelihood 1504 of a given bounding box score being incorrectly classified is determined by the ratio of the number of incorrectly labeled bounding boxes 1501 to the sum of the number of incorrectly labeled bounding boxes 1501 and the number of correctly labeled bounding boxes 1502. Fig.15 Observe that there is a peak 1505 in the possibilities 1504 at the best confidence threshold 1503 close to the pedestrian label. When selecting unlabeled bounding boxes for annotation, the active learning system 1400 will select bounding boxes with bounding box scores around the peak 1505.

[0144] Return to reference Fig.14 , the next step performed by the active learning system 1400 is to compute the spatial probability density over the set of bounding boxes for each label, sensor, and scale. Since most data samples are collected from the AV while driving, and since sensors (e.g., cameras, LiDAR, radar) are typically fixed in position, the spatial probability density can be used to identify anomalous bounding boxes that would not be suitable for annotation.

[0145] In an embodiment, a Gaussian mixture model (GMM) is applied to a set of bounding boxes for each tag, sensor (e.g., camera), and scale. The GMM is parameterized by the bounding box size (width, height) and bounding box position (x, y) in the sample reference frame, where x, y can be the position coordinates of the upper left corner of the bounding box in the sample reference frame (e.g., image reference frame). In order to view the spatial density as a probability, the spatial density is normalized to be in the range of [0, 1] by dividing the spatial density by the maximum spatial density value encountered in the bounding box set. In an embodiment, the minimum value of the maximum spatial density value can be enforced (e.g., 1e-4) to avoid division by zero errors caused by normalization.

[0146] Reference again Fig.16 , which shows the spatial probability density of the GMM for the "bike_without_rider" label and camera B0 for a medium-sized bounding box scale (e.g., box factor 1.0). In this example, camera B0 faces backwards, which is why the spatial probability density is symmetric along the vertical axis. If the camera faces the side of the AV100, the spatial probability density may not be symmetric.

[0147] The spatial probability density shows where a bicycle without a cyclist may be found in the image. The drawing can be viewed as a probability distribution of bicycles without a cyclist in the image space. It can be observed from the drawing that most of the density is concentrated in the middle band 1600, and there are possible peaks extending into the lower left and lower right sides of the drawing. When selecting a bounding box for annotation, the active learning system 1400 will select a bounding box in the middle band 1600. The bounding boxes in the upper part of the drawing and the lower part of the drawing may be abnormal. For example, since the upper part of the image represents the sky (assuming the AV is horizontal), it is impossible to detect a bicycle in the sky. Therefore, the bounding boxes located outside the middle band 1600 are excluded from the annotation.

[0148] Reference again Fig.14 , the next step performed by the active learning system 1400 is for the overall score variance module 1404 to calculate the overall score variance. Fig.17As shown in , N machine learning (ML) models 1701a-1701n (e.g., 3 neural networks) are trained using training data (e.g., training images). In an embodiment, a single base neural network is trained, N copies of the base neural network are created, and each copy is fine-tuned using training data in different random orders. The "ensemble" of machine learning models (e.g., CNNs) generates N prediction sets, which are fed into a pairwise comparator module 1702. The pairwise comparator module 1702 calculates the pairwise consistency between the N prediction sets to generate N overall scores (e.g., a set of N mean average precision (mAP) values), where each prediction set is in turn the "ground truth" of the other prediction sets at the box level. The overall score variance generator 1703 then calculates the overall score variance, which is a measure of the difference between the N overall scores. When selecting a data sample (e.g., an image) for annotation, the active learning system 1400 selects the data sample with the highest overall score variance (highest uncertainty). In an alternative embodiment, a cross-modal ensemble is used, where a measure of consistency is made between object detections output by multiple different sensors (eg, radar, LiDAR, and cameras).

[0149] In an embodiment, temporal inconsistency cues are used to determine whether consecutive data sample frames (e.g., consecutive video frames) contain temporal noise (e.g., "flicker"). For example, an object may disappear and then reappear in consecutive video frames. An object tracking algorithm (e.g., contour tracking, kernel-based tracking, CNN-based methods) can be used to track objects across frames to detect "flicker". Data samples that cause flicker are selected for annotation.

[0150] In an embodiment, a digital map is used to determine if an object violates a map constraint. For example, a car detected inside a building violates a map constraint. Bounding boxes that violate map constraints are excluded from selection for annotation. Map constraints need not be hard constraints. In an embodiment, the likelihood of a pedestrian being on a road, sidewalk, etc. is statistically modeled.

[0151] Fig.18 A flow chart showing an active learning process for automatically selecting data samples for annotation is shown. Process 1800 may be performed by Fig.14 The active learning system shown in uses e.g. Figure 3 The AV computer system 300 shown in FIG.

[0152] The process 1800 begins by obtaining a data sample (1801) including a bounding box and a corresponding bounding box score. For example, the data sample can be obtained as shown in FIG. Figure 4 and Figure 5An object detector in a perception module of the described AV is generated. The object detector may be a neural network (e.g., a CNN). The object detector may take as input an image and / or a point cloud captured by a camera and LiDAR / radar, respectively. The object detector outputs an image or a point cloud having a bounding box around the detected object, a label for marking the object, and a bounding box score indicating the confidence of the object detector that the bounding box is correctly labeled.

[0153] Process 1800 continues to generate a distribution of bounding box scores (1802). For example, process 1800 can generate a distribution of bounding box scores with the best confidence threshold 1503 ( Fig.15 ) are identified as incorrectly labeled, and bounding boxes with bounding box scores to the right of the optimal confidence threshold 1503 are identified as correctly labeled. The number of incorrectly and correctly labeled bounding boxes can then be used to create a stacked histogram to approximate the distribution, as shown in reference Fig.15 A likelihood is computed for each bin of the stacked histogram and used to select a bounding box for annotation.

[0154] Process 1800 continues with generating a spatial probability density over a set of bounding boxes for each label, sensor, and scale based on box size and position (1803). For example, a GMM parameterized by box size and position in the image sample can be used to generate a spatial probability density for a set of bounding boxes. The spatial probability density is used to determine the abnormal bounding boxes for selection for annotation, such as reference. Fig.16 As described.

[0155] Process 1800 continues to generate an overall score variance (1804) based on the prediction sets output from multiple machine learning models. For example, N machine learning (ML) models (e.g., 3 CNNs) are trained using training data (e.g., training images). The "overall" generation of the machine learning model is fed into the N prediction sets in the pairwise comparator module. The pairwise comparator module calculates the pairwise consistency between the N prediction sets to generate N overall scores (e.g., a set of N mAP values), where each prediction set is in turn the "ground truth" of the other prediction sets at the box level. The overall score variance generator then calculates the overall score variance, which is a measure of the difference between the N overall scores. When selecting a bounding box for annotation, the active learning system will select the data sample with the highest variance. In an optional embodiment, a cross-modal ensemble is used, in which a measure of consistency is made between object detections output by multiple different sensors (e.g., radar, LiDAR, and cameras).

[0156] Process 1800 continues with selecting data samples for annotation using a particular combination of bounding box score distribution, spatial probability density, and overall score variance (1805). In an embodiment, the data samples for annotation are selected using overall score variance and the distribution of bounding box scores without using spatial probability density. In an embodiment, the data samples for annotation are selected using overall score variance and spatial probability density without using the distribution of bounding box scores. In an embodiment, the data samples for annotation are selected using distribution of bounding box scores and spatial probability density without using overall score variance.

[0157] In the previous description, embodiments of the present invention have been described with reference to many specific details, which may be different due to implementation. Therefore, the description and the accompanying drawings should be regarded as illustrative, rather than restrictive. The only and exclusive indication of the scope of the present invention, and the applicant's expectation that the content of the scope of the present invention is the literal and equivalent scope of the claims issued from this application in the specific form of issuing the claims, including any subsequent amendments. Any definition of the terms used to be included in such claims clearly set forth herein should be based on the meaning of such terms as used in the claims. In addition, when the term "also includes" is used in the previous description or the attached claims, the following of the phrase can be an additional step or entity, or a sub-step / sub-entity of the previously described step or entity.

Claims

1. An active learning method, comprising: Obtaining a set of data samples using one or more processors, wherein each data sample includes one or more bounding boxes, each bounding box containing a potential object or scene in the environment, each bounding box having a label and a bounding box score indicating a confidence that the label is correct; and Selecting, using the one or more processors, a subset of data samples for annotation based on bounding box prediction confidences determined using a probability distribution of bounding box scores and an overall score variance based on differences in overall scores calculated from a set of predictions output by a plurality of machine learning models.

2. The method according to claim 1, further comprising: The subset of the data samples for annotation is selected using the one or more processors based on the bounding box prediction confidence, the spatial probability density of the bounding box parameterized by bounding box size and location, and the overall score variance.

3. The method according to claim 1, wherein: The bounding box prediction also includes: For each tag: Generate a probability distribution of bounding box scores; determining a likelihood that a particular bounding box is incorrectly labeled based on the distribution; and The particular bounding box is selected for annotation or excluded from annotation based on the likelihood.

4. The method according to claim 3, wherein: The distribution is approximated by a histogram with bins representing the range of bounding box scores, and each bin is associated with a likelihood.

5. The method according to claim 4, wherein: For each interval, the likelihood is calculated according to the ratio of the number of incorrectly labeled bounding boxes assigned to the interval to the sum of the number of incorrectly labeled bounding boxes and the number of correctly labeled bounding boxes assigned to the interval.

6. The method according to claim 2, further comprising: For each tag, sensor and scale: The spatial probability density is determined using a Gaussian mixture model (GMM) on a set of bounding boxes for the label, sensor, and scale, where the GMM is parameterized by bounding box size and position.

7. The method according to claim 6, wherein: The spatial probability density of the tag is determined by dividing the spatial density of the tag by the maximum density value among all spatial density values ​​of the tag.

8. The method according to claim 1, further comprising: Processing the data sample through a plurality of different machine learning models to generate predicted labeled bounding boxes; Calculate an overall score for each pairwise comparison of predicted labeled bounding boxes, where each predicted labeled bounding box is the ground truth for comparison with other predicted labeled bounding boxes; as well as An overall score variance is calculated based on the overall score.

9. The method according to claim 8, wherein: The multiple different machine learning models include multiple different neural networks tuned by training data samples provided by different types of sensors.

10. The method according to claim 9, wherein: The different types of sensors include light detection and ranging (LiDAR), radar, and cameras.

11. The method according to claim 9, wherein: The plurality of different neural networks are trained on training data samples in different random orders.

12. The method according to claim 1, further comprising: detecting, by the one or more processors, temporal inconsistencies between consecutive data samples; At least one data sample of the consecutive data samples is selected for annotation based on detecting the temporal inconsistency.

13. The method according to claim 1, further comprising: detecting, by the one or more processors, errors associated with the bounding box using the map constraints; as well as The bounding box is excluded from annotation based on detecting the error.

14. An active learning system comprising: a storage device comprising a data sample; one or more processors; a memory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform: Obtaining a set of data samples, wherein each data sample includes one or more bounding boxes, each bounding box contains a potential object or scene in the environment, and each bounding box has a label and a bounding box score indicating a confidence that the label is correct; as well as A subset of data samples is selected for annotation based on bounding box prediction confidence, which is determined using a probability distribution of bounding box scores, and an overall score variance, which is based on the difference in overall scores calculated from a set of predictions output by multiple machine learning models.

15. The system of claim 14, further comprising: The subset of data samples for annotation is selected based on the bounding box prediction confidence, the spatial probability density of the bounding box parameterized by bounding box size and position, and the overall score variance.

16. The system of claim 14, wherein: The bounding box prediction also includes: For each tag: Generate a probability distribution of bounding box scores; and determining a likelihood that a particular bounding box is incorrectly labeled based on the distribution; and The particular bounding box is selected for annotation or excluded from annotation based on the likelihood.

17. The system of claim 16, wherein: The distribution is approximated by a histogram with bins representing the range of bounding box scores, and each bin is associated with a likelihood.

18. The system of claim 17, wherein: For each interval, the likelihood is calculated based on the ratio of the number of incorrectly labeled bounding boxes assigned to the interval to the sum of the number of incorrectly labeled bounding boxes and the number of correctly labeled bounding boxes assigned to the interval.

19. The system of claim 15, further comprising: For each tag, sensor and scale: The spatial probability density is determined using a Gaussian mixture model (GMM) on a set of bounding boxes for the label, sensor, and scale, where the GMM is parameterized by bounding box size and position.

20. The system of claim 19, wherein: The spatial probability density of the tag is determined by dividing the spatial density of the tag by the maximum density value among all spatial density values ​​of the tag.

21. The system of claim 14, further comprising: Processing the data sample through a plurality of different machine learning models to generate predicted labeled bounding boxes; Calculate an overall score for each pairwise comparison of predicted labeled bounding boxes, where each predicted labeled bounding box is the ground truth for comparison with other predicted labeled bounding boxes; as well as An overall score variance is calculated based on the overall score.

22. The system of claim 21, wherein: The multiple different machine learning models include multiple different neural networks tuned by training data samples provided by different types of sensors.

23. The system of claim 22, wherein: The different types of sensors include light detection and ranging (LiDAR), radar, and cameras.

24. The system of claim 22, wherein: The plurality of different neural networks are trained on training data samples in different random orders.

25. The system of claim 14, further comprising: Detecting temporal inconsistencies between consecutive data samples; At least one data sample of the consecutive data samples is selected for annotation based on detecting the temporal inconsistency.

26. The system of claim 14, further comprising: Use map constraints to detect errors associated with bounding boxes; as well as The bounding box is excluded from annotation based on detecting the error.

27. A non-transitory computer-readable storage medium having instructions stored thereon, which when executed by at least one processor cause the at least one processor to perform the method according to any one of claims 1 to 13.

28. A computer program product comprising instructions which, when executed by at least one processor, cause the at least one processor to perform the method according to any one of claims 1 to 13.

Images