A data processing method and system for following a target positioning and motion control

Abstract

The application discloses a data processing method and system for following target positioning and motion control, and belongs to the technical field of data processing. The technical problem that the fusion of visual data and ultrasonic data is mostly simple superposition, the reliability difference of the two types of data in different scenes is not considered, and the precision of the fused distance data is insufficient in the field of data processing for following target positioning and motion control is solved. The method comprises the following steps: S1, sensor arrangement and data acquisition; S2, visual target detection based on NPU heterogeneous calculation; S3, monocular visual ranging and target locking; S4, target ranging based on ultrasonic waves; S5, longitudinal smoothing control based on a time window; and S6, lateral steering offset verification and posture alignment.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a data processing method and system for target positioning and motion control. Background Technology

[0002] In fields such as intelligent following devices and unmanned systems, precise positioning and smooth motion control of the target being followed are core technological requirements. The accuracy and real-time performance of data processing directly determine the following effect and operational safety of the device. Currently, existing data processing methods for target positioning and motion control mostly rely on a single sensor to collect and process data, which has significant limitations.

[0003] The single-vision sensor approach is susceptible to factors such as lighting and environmental texture, resulting in insufficient robustness in target recognition. Furthermore, monocular vision ranging inherently suffers from scale ambiguity, hindering accurate ranging. Additionally, the computational demands of complex deep learning models are substantial, making it difficult to meet real-time requirements on embedded platforms. The single-ultrasonic sensor approach, on the other hand, suffers from measurement blind spots, susceptibility to environmental interference leading to echo distortion, and cannot achieve accurate target recognition and locking, making it unsuitable for dynamic tracking scenarios.

[0004] Furthermore, in existing technologies, the fusion of visual and ultrasonic data is mostly a simple superposition, without considering the reliability differences between the two types of data in different scenarios, resulting in insufficient accuracy of the fused distance data. In terms of motion control, there is a lack of effective smoothing strategies and offset verification mechanisms, which can easily lead to problems such as sudden speed changes and steering deviations, affecting the stability and accuracy of following.

[0005] Therefore, how to integrate the advantages of multi-sensor data to achieve efficient target recognition and accurate distance measurement, while optimizing motion control strategies and improving the real-time performance, stability, and accuracy of the following process, has become a pressing technical problem to be solved in the field of target positioning and motion control data processing. Summary of the Invention

[0006] To address the current technical problem in target tracking and motion control data processing where the fusion of visual and ultrasonic data is often a simple superposition without considering the reliability differences between the two types of data in different scenarios, resulting in insufficient accuracy of the fused distance data, this invention proposes a data processing method and system for target tracking and motion control. The method and system of this invention are applicable to subjects to be controlled at speeds up to 25 km / h.

[0007] The method includes the following steps: S1. Sensor deployment and data acquisition: A camera and an ultrasonic sensor are deployed on the main body that needs to be positioned and controlled to follow the target; the camera collects environmental data in front of the main body, and the ultrasonic sensor collects ultrasonic data in front of the main body. S2. Visual target detection based on NPU heterogeneous computing: The video data stream is acquired through the camera, and the current video frame is read; for each video frame, the target object is identified using a neural network processor (NPU); S3. Monocular Vision Ranging and Target Locking: Calculate the distance to the identified target object. When multiple target objects are identified, lock the smallest one. The target object corresponding to the value is taken as the unique follower target; S4. Ultrasonic Target Ranging: Based on the locked target, a confidence window is set. Within the confidence window, high-gain ultrasonic acquisition is enabled, and the final target distance is output. ; S5. Time-Window-Based Vertical Smoothing Control: A biomimetic smoothing strategy combining a ladder state machine and a time window is employed. Output the strategy for controlling the speed of the main body; S6. Lateral Rotation Offset Verification and Attitude Alignment: Extract the center offset of the target object within the visual target detection box. According to the offset Output the strategy for steering control of the main body.

[0008] Furthermore, before using the Neural Processing Unit (NPU) for target object recognition, the video frames are preprocessed. Specifically, the Letterbox algorithm is used to scale and fill the original frames to the size required by the detection model. The NPU uses the YOLO model for target object recognition.

[0009] further, ,in, The preset target average body width constant, The equivalent focal length of the calibrated camera. This is a hardware offset constant introduced to eliminate fixed errors caused by the camera's installation position. The detected target bounding box is ,in, These are the x-coordinates of the top left corner, y-coordinates of the top left corner, the x-coordinates of the bottom right corner, and the y-coordinates of the bottom right corner of the target bounding box, respectively.

[0010] Furthermore, the confidence window is , , ,in, For the speed of sound, ,in, For distance tolerance, This indicates the tailing time of the ultrasonic transducer.

[0011] further, ,in, This indicates the distance to the target object detected by an ultrasonic sensor. It is a dynamic weighting factor.

[0012] Furthermore, dynamic weighting factors Based on the confidence level of visual detection Dynamic adjustment ,in, , and These are the weighting coefficients; For the confidence term of the target object recognition model, For morphological geometric stability, This is a spatiotemporal consistency term.

[0013] Furthermore, dynamic weighting factors Based on the confidence level of visual detection The dynamic adjustment is specifically as follows: ,in, The preset trust threshold, This is the sensitivity coefficient.

[0014] Furthermore, combined with When outputting a strategy for speed control of the subject, two strategies are included: Strategy 1, based on Define the target speed gear of the main body : ; represent ; These represent different distances, increasing sequentially. The distance is measured in meters; the speeds corresponding to gears 0-6 increase sequentially. Strategy 2: Set a time window limit: Let the current time be... The last gear change time was Only when: Acceleration or deceleration commands are only allowed to be executed after a certain number of seconds, and the value of h is set to within 1.

[0015] Furthermore, based on the offset The specific strategy for steering control of the main body is as follows: setting a steering zone threshold. ,when At that time, it was determined that the main body needed to be adjusted in direction. Determine to turn left. At that time, a right turn is determined.

[0016] The present invention also provides a data processing system for target positioning and motion control, comprising: The unit for sensor deployment and data acquisition: Cameras and ultrasonic sensors are deployed on the main body that needs to perform target tracking, positioning, and motion control; the cameras collect environmental data in front of the main body, and the ultrasonic sensors collect ultrasonic data in front of the main body; The unit that performs visual target detection based on NPU heterogeneous computing: acquires video data streams through a camera and reads the current video frame; for each video frame, it uses a neural network processor (NPU) to identify the target object; The unit that performs monocular vision ranging and target locking: For the identified target object, calculates the distance to the target object. When multiple target objects are identified, lock the smallest one. The target object corresponding to the value is taken as the unique follower target; The unit for ultrasonic-based target ranging: Based on the locked target, a confidence window is set, and high-gain ultrasonic acquisition is enabled within the confidence window to output the final target distance. ; The unit implements time-window-based longitudinal smoothing control: It employs a biomimetic smoothing strategy combining a ladder state machine with a time window. Output the strategy for controlling the speed of the main body; The unit that performs lateral steering offset verification and pose alignment: extracts the center offset of the target object within the visual target detection box. According to the offset Output the strategy for steering control of the main body.

[0017] The beneficial effects of the method described in this invention are as follows: By using visual and ultrasonic sensors to collect data collaboratively, instead of simply superimposing the two types of data, a confidence window is set based on the locked target. Within the window, high-gain ultrasonic acquisition is enabled, and a dynamic weighting factor is introduced to dynamically adjust the fusion weight of visual ranging data and ultrasonic ranging data according to the confidence level of visual detection. This fully considers the reliability differences between the two types of data in different scenarios, which greatly improves the accuracy of the final target distance after fusion and effectively avoids the limitations of single-sensor ranging.

[0018] In addition, NPU heterogeneous computing is used to achieve efficient visual target detection, improving the real-time performance and robustness of recognition; the uniqueness of the following target is ensured by combining monocular visual ranging with target locking strategy; and the longitudinal smoothing control and lateral steering offset verification by combining a ladder state machine with time window are further improved to enhance the stability and accuracy of following motion. Overall, the comprehensive performance of following target positioning and motion control is optimized to meet the dynamic following needs of multiple scenarios. Attached Figure Description

[0019] Figure 1 This is a general flowchart of the method described in the embodiments of the present invention; Figure 2 This is a diagram of the device architecture used in the method described in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the visual detection and ranging principle in an embodiment of the present invention; Figure 4 This is a logic diagram for longitudinal stepped speed control in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0021] Example 1 This embodiment provides a data processing method for target localization and motion control, such as... Figure 1 As shown, it includes the following steps: S1. Sensor deployment and data acquisition: A camera and an ultrasonic sensor are deployed on the main body that needs to be positioned and controlled to follow the target; the camera collects environmental data in front of the main body, and the ultrasonic sensor collects ultrasonic data in front of the main body. S2. Visual target detection based on NPU heterogeneous computing: The video data stream is acquired through the camera, and the current video frame is read; for each video frame, the target object is identified using a neural network processor (NPU); Before using a neural network processor (NPU) for target object recognition, the video frames are preprocessed. Specifically, the Letterbox algorithm is used to scale and fill the original frames to the size required by the detection model. The NPU uses the YOLO model for target object recognition.

[0022] S3. Monocular Vision Ranging and Target Locking: Calculate the distance to the identified target object. When multiple target objects are identified, lock the smallest one. The target object corresponding to the value is taken as the unique follower target; ,in, The preset target average body width constant, The equivalent focal length of the calibrated camera. This is a hardware offset constant introduced to eliminate fixed errors caused by the camera's installation position. The detected target bounding box is ,in, These are the x-coordinates of the top left corner, y-coordinates of the top left corner, the x-coordinates of the bottom right corner, and the y-coordinates of the bottom right corner of the target bounding box, respectively.

[0023] S4. Ultrasonic Target Ranging: Based on the locked target, a confidence window is set. Within the confidence window, high-gain ultrasonic acquisition is enabled, and the final target distance is output. ; Confidence window is , , ,in, For the speed of sound, ,in, For distance tolerance, This indicates the tailing time of the ultrasonic transducer.

[0024] ,in, This indicates the distance to the target object detected by an ultrasonic sensor. It is a dynamic weighting factor.

[0025] Dynamic weighting factor Based on the confidence level of visual detection Dynamic adjustment ,in, , and These are the weighting coefficients; For the confidence term of the target object recognition model, For morphological geometric stability, This is a spatiotemporal consistency term.

[0026] ,in, The preset trust threshold, This is the sensitivity coefficient.

[0027] S5. Time-Window-Based Vertical Smoothing Control: A biomimetic smoothing strategy combining a ladder state machine and a time window is employed. Output the strategy for controlling the speed of the main body; S6. Lateral Rotation Offset Verification and Attitude Alignment: Extract the center offset of the target object within the visual target detection box. According to the offset Output the strategy for steering control of the main body.

[0028] Based on offset The specific strategy for steering control of the main body is as follows: setting a steering zone threshold. ,when At that time, it was determined that the main body needed to be adjusted in direction. Determine to turn left. At that time, a right turn is determined.

[0029] Example 2 This embodiment takes a bicycle as the subject to be controlled and a pedestrian as the tracking target as an example to further explain the method in step 1.

[0030] S1. Sensor placement and data acquisition: like Figure 2 As shown, the device nodes are traversed through the V4L2 interface to dynamically detect and bind camera indices for valid image streams. To prevent camera reading from blocking the main control logic, a video stream reading queue based on a daemon thread is constructed. The camera resolution is initialized to 640×480 (or 1280×720), and the maximum number of frames in the internal buffer of the camera / video stream is set to 1 to ensure that each read frame is the latest frame, eliminating the accumulation of video latency.

[0031] S2. Visual object detection based on NPU heterogeneous computing: After extracting the current video frame, the system uses the RKNN-Lite framework to offload the computation task to the Rockchip NPU's NPU_CORE_0.

[0032] Image preprocessing (Letterbox): To maintain the image's aspect ratio, the Letterbox algorithm is used to scale and fill the original frame to the required 640×640 size. Scaling ratio The calculation formula is: ,in, and These represent the height and width of the input image required by the object detection model (both are 640 pixels in this embodiment); and These represent the height and width of the original video frame actually captured by the camera, respectively.

[0033] Use pure black RGB(0,0,0) to fill the insufficient area with a constant border.

[0034] Model inference and post-processing: The preprocessed tensor is input into the YOLO model to extract predicted bounding boxes with a confidence score greater than 0.25 and a class index of 0 (i.e., human). Then, non-maximum suppression (NMS) is used to remove overlapping boxes, with the intersection-over-union (IoU) threshold set to 0.45.

[0035] S3. Monocular visual ranging and target locking: For the detected pedestrian bounding boxes ,in, These are the x-coordinates of the top left corner, y-coordinates of the top left corner, the x-coordinates of the bottom right corner, and the y-coordinates of the bottom right corner of the target bounding box, respectively.

[0036] Calculate its pixel width in the image .like Figure 3 As shown, a physical distance estimation formula is established based on the pinhole camera model and the principle of similar triangles. To eliminate the fixed error caused by the camera installation position, a hardware offset constant is introduced. Target distance The calculation formula is: ; in, The average shoulder width / body width of pedestrians is a preset constant (0.5m in this embodiment). The equivalent focal length of the camera after calibration (880 in this embodiment). The physical compensation distance from the camera to the axle of the bicycle's front wheel is 0.4m in this embodiment.

[0037] When multiple people are present in the frame, the system presses... Sort by size from smallest to largest, automatically lock the nearest pedestrian as the sole primary follower, and extract their distance. and center point coordinates .

[0038] S4. Ultrasonic-based target ranging: By using the visual recognition results of the NPU as "seed information", a "confidence window" is dynamically defined on the echo time axis of the ultrasound, and environmental clutter (such as utility poles, walls, and other pedestrians) outside the window is eliminated, thereby achieving acoustic focusing on a specific target.

[0039] Dynamic "time window" opening algorithm: The system according to Calculate the expected arrival time of the echo To tolerate visual errors, in The front and back opening width is Time window: Lower limit of the window: ; Window limit: ; ,in, For distance tolerance, This indicates the tail time of the ultrasonic transducer, in this embodiment. Take 1ms.

[0040] Distance tolerance is the maximum displacement deviation caused by the target object (pedestrian) moving back and forth or inaccurate edge extraction at the current sampling rate under monocular vision.

[0041] Ultrasonic transducer tail time: The aftershock time of an ultrasonic transducer due to mechanical inertia after excitation stops. This time must be included in the window to ensure complete signal acquisition. Typical value: .

[0042] Logic processing: The ultrasonic receiving circuit only... High-gain acquisition is enabled during the time period, and echo signals outside the window are identified as background noise and suppressed.

[0043] Enabling high-gain acquisition refers to dynamically amplifying the echo signal at the receiving end in the time domain. In practice, the ultrasonic sensor receiver is equipped with a programmable gain amplifier (PGA). Outside of the acquisition window, the ultrasonic sensor maintains an extremely low base gain to suppress ambient background noise (such as road surface reflection interference); when the visually guided gating window is reached... Within this timeframe, the main control unit of the ultrasonic sensor instantly sends a command, significantly increasing the voltage amplification factor of the receiving circuit (i.e., enabling high gain). This mechanism acts like an "acoustic spotlight," precisely and specifically amplifying weak target reflection signals at a specific depth without amplifying global environmental noise, significantly improving the signal-to-noise ratio of the target characteristic waveform during analog-to-digital converter (ADC) sampling.

[0044] Audiovisual dual-modal weighted fusion model: To further improve locking accuracy, the system introduces a dynamic weighting factor. Based on the confidence level of visual detection Dynamically adjust ranging weights: ; Scene adaptation: When a pedestrian is briefly obscured, When decreasing, increase (Relying on visual prediction and inertial navigation); when lighting is good and visual features are obvious, reduce It primarily uses ultrasonic echo data with higher precision.

[0045] ; (Model Confidence Term): The category probability (Objectness Score) directly output by the YOLO model used for target object recognition, representing the degree to which "this looks like a person".

[0046] (Morphological geometric stability term): Based on the pedestrian's prior aspect ratio verification. ,in, , To detect the height and width of the frame, This is the standard human body proportion constant, approximately 2.5 to 3.0. This is the proportional attenuation coefficient, used to control the algorithm's sensitivity to deviations. If... The larger the bounding box, the more stringent the shape requirements become in the algorithm. Even slight widening or flattening of the detection box (e.g., misidentifying a utility pole or trash can) will result in a poorly executed algorithm. It will instantly approach ;if A smaller setting increases the system's tolerance, allowing pedestrians a certain degree of bending or arm swinging. If the scale is distorted, the score drops sharply, effectively eliminating false recognition of background clutter.

[0047] (Spatiotemporal consistency term): Measures the logical displacement of the target across consecutive frames. , This represents the bounding box of the target detected in the current frame. This represents the target bounding box predicted based on the state of the previous frame. This is the intersection-over-union ratio (IoU) between the predicted bounding box and the current detected bounding box. It is calculated using the IoU between the predicted bounding box and the current detected bounding box, combined with the Euclidean distance offset. Evaluation. If the target shifts instantaneously, it is considered a false trigger, and the score is reduced. The Euclidean distance offset is the linear distance in space between the center point of the detection box in the current frame and the center point of the prediction box in the previous frame.

[0048] As a weighting coefficient for visual distance, its adjustment logic follows the "trust and authorization" principle: High confidence state ( When the visual system clearly locks onto a target, its shape is stable, and its movement is smooth, the system has a high degree of trust in the "visual guide seed." At this point, the [visual guide seed] should be reduced. (make (Increase), thereby shifting the system focus to the ultrasonic echo guided by a visually precise window, to obtain centimeter-level ranging accuracy.

[0049] Low confidence state ( When the target is partially occluded, experiences drastic changes in lighting, or suffers morphological distortion, the visual seed becomes unreliable. As the dynamics increase, the system enters either "visual prediction maintenance" or "safe and conservative mode," increasing the reference weight of the visual historical trajectory to avoid drastic changes in ranging data caused by incorrect ultrasonic windowing.

[0050] In order to avoid The step jump causes a jerk in motor control, so a sigmoid mapping function is used: ;in The preset trust threshold, This is the sensitivity coefficient. (This ensures a smooth transition in weights when the visual environment changes from "excellent" to "normal".)

[0051] S5. Vertical smoothing control based on time window: When the target is visually locked, longitudinal speed control is executed. Traditional PID continuous control is abandoned, and a biomimetic smoothing strategy combining a stepped state machine and a time window is adopted to adapt to the discrete command characteristics of the bicycle motor, such as... Figure 4 The diagram shown is a longitudinal stepped speed control logic diagram in this embodiment.

[0052] Distance and Gear Mapping: Define the target speed gear For piecewise functions: ; The speed setting for each gear (1-6) must meet certain conditions. Specifically, the speed is 3.6 km / h in gear 0, and the speed increases by 1 km / h for each subsequent gear, with the maximum speed not exceeding 25 km / h.

[0053] Time window limiting: To prevent motor overload or severe vehicle body shaking due to sudden command changes, an acceleration time control window is introduced. Let the current time be... The last time the speed changed was Only when: Acceleration / deceleration commands are only allowed when [the specified conditions are met]. Then send to the serial port Each 'w' acceleration command increases the speed by one gear; if If so, a deceleration command 'p' is sent, thereby achieving a smooth step transition in speed. To determine the target speed gear, the system calculates the desired gear based on the final distance calculated using a preset piecewise function. For example, if the target pedestrian is 2.5 meters away, a table would indicate that "gear 4" should be engaged. Current speed gear: The actual gear that the bicycle's bottom motor controller is currently operating at.

[0054] S6. Lateral steering offset verification and attitude alignment: Extract the screen center offset of the target (For an image with a width of 640, ).

[0055] Set steering dead zone threshold Pixel. When When necessary, the handlebar angle needs adjustment. A steering time window limiting mechanism is also introduced (0.3s in this embodiment) to avoid high-frequency oscillations in the servo or steering motor. Send the left turn command 'l'; if Send the right turn command 'r'.

Claims

1. A data processing method for target localization and motion control, characterized in that, The method includes the following steps: S1. Sensor deployment and data acquisition: A camera and an ultrasonic sensor are deployed on the main body that needs to be positioned and controlled to follow the target; the camera collects environmental data in front of the main body, and the ultrasonic sensor collects ultrasonic data in front of the main body. S2. Visual target detection based on NPU heterogeneous computing: The video data stream is acquired through the camera, and the current video frame is read; for each video frame, the target object is identified using a neural network processor (NPU); S3. Monocular Vision Ranging and Target Locking: Calculate the distance to the identified target object. When multiple target objects are identified, lock the smallest one. The target object corresponding to the value is taken as the unique follower target; S4. Ultrasonic Target Ranging: Based on the locked target, a confidence window is set. Within the confidence window, high-gain ultrasonic acquisition is enabled, and the final target distance is output. ; S5. Time-Window-Based Vertical Smoothing Control: A biomimetic smoothing strategy combining a ladder state machine and a time window is employed. Output the strategy for controlling the speed of the main body; S6. Lateral Rotation Offset Verification and Attitude Alignment: Extract the center offset of the target object within the visual target detection box. According to the offset Output the strategy for steering control of the main body.

2. The data processing method for target positioning and motion control according to claim 1, characterized in that, Before using a neural network processor (NPU) for target object recognition, the video frames are preprocessed. Specifically, the Letterbox algorithm is used to scale and fill the original frames to the size required by the detection model. The NPU uses the YOLO model for target object recognition.

3. The data processing method for target positioning and motion control according to claim 2, characterized in that, ,in, The preset target average body width constant, The equivalent focal length of the calibrated camera. This is a hardware offset constant introduced to eliminate fixed errors caused by the camera's installation position. The detected target bounding box is ,in, These are the x-coordinates of the top left corner, y-coordinates of the top left corner, the x-coordinates of the bottom right corner, and the y-coordinates of the bottom right corner of the target bounding box, respectively.

4. The data processing method for target positioning and motion control according to claim 3, characterized in that, Confidence window is , , ,in, For the speed of sound, ,in, For distance tolerance, This indicates the tailing time of the ultrasonic transducer.

5. The data processing method for target positioning and motion control according to claim 4, characterized in that, ,in, This indicates the distance to the target object detected by an ultrasonic sensor. It is a dynamic weighting factor.

6. The data processing method for target positioning and motion control according to claim 5, characterized in that, Dynamic weighting factor Based on the confidence level of visual detection Dynamic adjustment ,in, , and These are the weighting coefficients; For the confidence term of the target object recognition model, For morphological geometric stability, This is a spatiotemporal consistency term.

7. The data processing method for target positioning and motion control according to claim 6, characterized in that, Dynamic weighting factor Based on the confidence level of visual detection The dynamic adjustment is specifically as follows: ,in, The preset trust threshold, This is the sensitivity coefficient.

8. The data processing method for target positioning and motion control according to claim 7, characterized in that, Combination When outputting a strategy for speed control of the subject, two strategies are included: Strategy 1, based on Define the target speed gear of the main body : ； represent ; These represent different distances, increasing sequentially. The distance is measured in meters; the speeds corresponding to gears 0-6 increase sequentially. Strategy 2: Set a time window limit: Let the current time be... The last gear change time was Only when: Acceleration or deceleration commands are only allowed to be executed after a certain number of seconds, and the value of h is set to within 1.

9. The data processing method for target positioning and motion control according to claim 8, characterized in that, Based on offset The specific strategy for steering control of the main body is as follows: setting a steering zone threshold. ,when At that time, it was determined that the main body needed to be adjusted in direction. Determine to turn left. At that time, a right turn is determined.

10. A data processing system for target positioning and motion control, characterized in that, include: The unit for sensor deployment and data acquisition: Cameras and ultrasonic sensors are deployed on the main body that needs to perform target tracking, positioning, and motion control; the cameras collect environmental data in front of the main body, and the ultrasonic sensors collect ultrasonic data in front of the main body; The unit that performs visual target detection based on NPU heterogeneous computing: acquires video data streams through a camera and reads the current video frame; for each video frame, it uses a neural network processor (NPU) to identify the target object; The unit that performs monocular vision ranging and target locking: For the identified target object, calculates the distance to the target object. When multiple target objects are identified, lock the smallest one. The target object corresponding to the value is taken as the unique follower target; The unit for ultrasonic-based target ranging: Based on the locked target, a confidence window is set, and high-gain ultrasonic acquisition is enabled within the confidence window to output the final target distance. ; The unit implements time-window-based longitudinal smoothing control: It employs a biomimetic smoothing strategy combining a ladder state machine with a time window. Output the strategy for controlling the speed of the main body; The unit that performs lateral steering offset verification and pose alignment: extracts the center offset of the target object within the visual target detection box. According to the offset Output the strategy for steering control of the main body.

Images