A method and device for intelligent cone barrel action recognition based on multi-level feature discrimination

Abstract

The application discloses a kind of intelligent cone bucket action recognition method and device based on multi-level feature discrimination, comprising the following steps: step S1: when the instantaneous change amount of acceleration exceeds a first dynamic threshold Th1, trigger a first wake-up;Step S2: start the high-frequency sampling mode of inertial sensor module, continuously collect acceleration data and gyroscope data in the preset time window;Step S3: the multi-dimensional feature vector F is input into the preset multi-level discrimination model;Step S4: according to the action classification result, determine whether position reporting and event reporting are needed;Step S5: start GNSS module to obtain the accurate latitude and longitude coordinates of cone bucket;Step S6: start communication module, upload to cloud server;Step S7: the first dynamic threshold Th1 is adaptively updated according to the following formula.A closed-loop technical solution from "low-power perception" to "accurate identification" to "intelligent reporting and adaptation" is constructed.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and the Internet of Things, specifically to an intelligent cone action recognition method and device based on multi-level feature discrimination. Background Technology

[0002] Traffic cones (commonly known as "ice cream cones") are the most commonly used safety warning and traffic separation facilities in scenarios such as highway construction, accident handling, and temporary traffic control. With the development of intelligent transportation and vehicle-road cooperative technologies, the industry's demand for cones has evolved from passive physical warnings to intelligent capabilities such as real-time perception, location reporting, and status feedback, so as to obtain information on road construction and abnormal road conditions in real time, thereby improving traffic efficiency and the safety of workers.

[0003] Existing smart cones or cone status sensing devices mainly fall into the following categories: (1) Cone state detection device based on a single sensor: It mainly detects the tilting state of the cone through a single tilt angle sensor or acceleration sensor and reports it through a wireless communication module. This type of device has a single function and can only identify two states: "tilted" or "not tilted". It cannot distinguish between different types of actions such as the cone being moved by hand, hit by a vehicle, or blown over by the wind.

[0004] (2) Smart cones based on continuous location reporting: These devices work continuously through a GNSS module and periodically report location information, and trigger reporting by combining simple motion detection. Due to the high power consumption of the GNSS module and the short battery life, these devices can usually only last for a few days to a few weeks, which is difficult to meet the needs of long-term field deployment of cones.

[0005] (3) Simple IoT cone based on motion wake-up: After detecting a motion event by an accelerometer, the main control chip is woken up and the positioning and communication module is started to report. Although this type of device reduces the average power consumption, it has two prominent problems: First, the granularity of motion event classification is coarse, and it cannot distinguish between different behaviors such as "placement", "transportation", and "impact"; Second, continuous triggering during transportation leads to frequent wake-up of the positioning and communication module, resulting in a large number of invalid reports and energy waste.

[0006] It is evident that the main problems with existing technologies include: insufficient action recognition capabilities, limited to simple binary classification (moving / stationary or tilting / standing); high false alarm and false negative rates, with fixed thresholds that cannot be adaptively adjusted; inefficient power management and short battery life; and low-value reported data, failing to support advanced application scenarios. Therefore, there is an urgent need for an intelligent cone action recognition method and device capable of accurately recognizing various typical cone actions while possessing low power consumption characteristics. Summary of the Invention

[0007] In view of this, the purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent cone action recognition method based on multi-level feature discrimination.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent cone action recognition method based on multi-level feature discrimination, comprising the following steps: Step S1: Monitor the acceleration signal of the cone in low-power mode using the inertial sensor module. When the instantaneous change in the resultant acceleration exceeds the first-level dynamic threshold Th1, trigger the first-level wake-up and proceed to step S2. Step S2: Activate the high-frequency sampling mode of the inertial sensor module, continuously collect acceleration data and gyroscope data within a preset time window, and extract a multi-dimensional feature vector F; the feature vector F includes time domain features, frequency domain features, attitude domain features and displacement features; Step S3: Input the multidimensional feature vector F into a preset multi-level discrimination model, calculate the confidence score of each type of action, and output the action classification result; the action classification includes at least six categories: deployment and placement, retraction and removal, vehicle collision, illegal movement, natural dumping and abnormal water immersion. Step S4: Determine whether location reporting and event reporting are required based on the action classification results. Only proceed to the next step if the determination result is that the event needs to be reported; otherwise, return to step S1 to continue monitoring. Step S5: Activate the GNSS module to obtain the precise latitude and longitude coordinates of the cone; Step S6: Start the communication module and upload the event type, location data, and device status information to the cloud server. Step S7: Adaptively update the first-level dynamic threshold Th1 according to the following formula: Th1=max(0.2g, Δa) 95 ×1.5), where Δa 95 It represents the 95th percentile of the instantaneous change in the combined acceleration over the most recent hour.

[0009] Optimally, in step S2, the time-domain features include the peak resultant acceleration A_peak, the mean resultant acceleration A_mean, the peak-to-average power ratio (PMR) of the resultant acceleration, the vibration attenuation coefficient Decay, the vibration stability σ_end, and the motion duration T_motion; the frequency-domain features include the discrete Fourier transform X(k), the power spectral density P(k), the main peak frequency F_main, the high-frequency energy proportion E_h, and the spectral entropy H_spec; the attitude-domain features include the cumulative change in attitude angle Δθ, the peak rate of attitude angle change ω_peak, the final attitude angle θ_end, the final attitude stability σ_θ, and the attitude change path length L_θ.

[0010] Furthermore, in step S2, the displacement feature includes at least the fused displacement Δd_hybrid, which is obtained by weighting the GNSS displacement and the displacement calculated by inertial dead reckoning. The formula for calculating the displacement characteristic Δd_hybrid is as follows: Δd_hybrid = K_g · Δd_gnss + (1-K_g) ·Δd_ins; K_g = HDOP_norm × Sats_norm; In the formula: K_g is the GNSS confidence weight, Δd_gnss is the displacement based on GNSS, Δd_ins is the displacement based on inertial dead reckoning, HDOP_norm=1 / max(HDOP,1), Sats_norm=min(Sats / 6,1).

[0011] Furthermore, in step S3, for the two types of actions, "illegal movement" and "removal", a three-dimensional boundary discrimination criterion including termination posture stability, displacement distance, and motion continuity is used for refined classification.

[0012] Furthermore, the termination attitude stability judgment D1 is based on the attitude angle θ(t) and its standard deviation σ_θ(t) within the observation window T_obs after the action stops; The displacement distance discrimination D2 is based on comparing the fused displacement Δd_hybrid with the preset maximum allowable displacement distance D_max; The motion continuity determination D3 is based on the number of motion phases N_phases and the longest single-phase motion duration T_phase_i during the motion process; The comprehensive score S_illegal is obtained by weighted fusion, which is calculated as w1·D1 + w2·D2 + w3·D3. When S_illegal is greater than or equal to the preset threshold τ_min, it is judged as "illegal movement". Otherwise, it is judged as "reclaim and move" or enters the secondary confirmation window.

[0013] Furthermore, step S3 also includes a secondary confirmation window mechanism: When the confidence scores of all action classifications D_k(F) < τ_min, the system enters a pending confirmation state with a duration of T_conf = 30s. After reaching T_conf, the final state attitude is determined: if |θ_end|<30° and σ_θ<5°, it is determined as an illegal movement; If |θ_end|>60° or the movement continues, it is determined that the object should be retrieved and moved away. All other cases are classified as natural spillage.

[0014] Another object of the present invention is to provide an intelligent cone action recognition device based on multi-level feature discrimination, comprising: Low-power microcontrollers; An inertial sensor module, which is electrically connected to the low-power microcontroller, is used to collect the acceleration and angular velocity data of the cone. A GNSS module, electrically connected to the low-power microcontroller, is used to acquire the latitude and longitude coordinates of the cone. A communication module, which is electrically connected to the low-power microcontroller, is used to upload data to a cloud server. A water immersion sensor, which is electrically connected to a low-power microcontroller; A power management module is electrically connected to the low-power microcontroller, the inertial sensor module, the GNSS module, the communication module, and the water immersion sensor. A lithium-ion battery, wherein the lithium-ion battery is electrically connected to the power management module; A housing for accommodating the aforementioned components; the water immersion sensor is located at the bottom of the housing; The low-power microcontroller is configured to execute the aforementioned intelligent cone action recognition method.

[0015] Ideally, the low-power microcontroller is equipped with a device lifecycle state machine, which includes at least the following states: The deployed status S1 indicates that the cones have been stably placed in the work area; The S5 status during transport indicates that the cone has been confirmed to be retrieved and moved, and the cone continues to move. The S6 status indicates that the device has been stationary in a non-deployment position for an extended period of time, and its return to the warehouse has been confirmed. The S7 indicates that the device has entered deep sleep mode after a long period of inactivity.

[0016] Ideally, the state machine reports the corresponding command word message during state transitions: When the process transitions from the moving state S5 to the recovered state S6, a command word "collected" message is reported. When transitioning from the recovered state S6 to the offline state S7, an offline command message is reported.

[0017] Ideally, the inertial sensor module includes an accelerometer and a gyroscope, which are rigidly fixed on the central axis inside the housing to ensure that its measurement axis is consistent with the geometric axis of the cone; the housing has a protection rating of IP67 or higher.

[0018] Compared with existing technologies, the present invention has the following beneficial effects: The intelligent cone action recognition method based on multi-level feature discrimination of the present invention constructs a closed-loop technical solution from "low-power perception" to "accurate recognition" and then to "intelligent reporting and adaptation". The technical effects it brings are systematic and synergistic, specifically reflected in the following aspects: (1) Achieving ultra-low power consumption and long battery life: This enables the device to be powered by lithium-ion batteries and to work continuously in the field for more than 3 years, which greatly reduces maintenance costs and meets the stringent requirements for long-term deployment of intelligent transportation infrastructure. (2) Improve the accuracy of action recognition and enrich the types of recognition: It can accurately distinguish at least six types of cone actions with clear semantics. The recognition accuracy can reach more than 90% in the test environment, providing high-value event data for upper-layer applications; (3) Improve environmental adaptability and reduce false alarm / false alarm rate: It can automatically adapt to different deployment environments, such as quiet urban roads, high-vibration highways, and heavy machinery construction areas, significantly reducing false alarms and false alarms caused by changes in the environmental vibration baseline; (4) Provide high-value data to support in-depth applications: The data uploaded to the cloud is no longer a simple "location" or "dumping", but an "event type" with clear semantics (such as vehicle collision, illegal movement), which can directly provide Gaode / Baidu Maps with real-time traffic conditions, provide traffic management departments with construction area supervision, and provide property owners with data support for in-depth applications such as asset theft prevention.

[0019] (5) Real-time recognition on the device side reduces system dependence: The entire action recognition process is completed on the low-power microcontroller on the device itself, with recognition latency as low as seconds, and it does not depend on cloud computing power and network connection at all. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the intelligent cone action recognition device based on multi-level feature discrimination according to the present invention; Figure 2 This is a hardware system block diagram of the intelligent cone action recognition device based on multi-level feature discrimination according to the present invention; Figure 3 This is an overall flowchart of the intelligent cone action recognition method based on multi-level feature discrimination of the present invention; Figure 4 This is a schematic diagram of the state transitions of the four-level hierarchical wake-up architecture of the present invention; Figure 5 These are schematic diagrams of characteristic waveforms for six typical actions of this invention; Figure 6 This is a logic block diagram of the action classification and determination model of the present invention; Figure 7 This is a schematic diagram of the adaptive update mechanism for the environmental vibration baseline of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will now be described in detail.

[0022] The foregoing descriptions and other technical contents, features, and effects of this invention will be clearly presented in the following detailed description of a preferred embodiment with reference to the accompanying drawings. The directional terms used in the following embodiments, such as up, down, left, right, front, or back, are merely for reference to the accompanying drawings. Therefore, the directional terms used are for illustrative purposes and not for limiting the invention. Example 1

[0023] This embodiment provides an intelligent cone action recognition device based on multi-level feature discrimination, such as... Figure 1 and Figure 2 As shown, it mainly includes a low-power microcontroller 1, an inertial sensor module 2, a GNSS module 3 (i.e., a sub-meter level GNSS positioning module), a communication module 4 (i.e., a 4G Cat1 communication module), a water immersion sensor 5, a power management module 6, a lithium-ion battery 7, and a housing 8.

[0024] The housing 8 is used to house the above components (low-power microcontroller 1, inertial sensor module 2, GNSS module 3, communication module 4, water immersion sensor 5, power management module 6, and lithium-ion battery 7).

[0025] The low-power microcontroller 1 uses an ultra-low-power MCU (such as the STM32L series) with an ARM Cortex-M0+ core, featuring a current consumption of less than 2μA in deep sleep mode and supporting fast wake-up via external interrupts. The inertial sensor module 2 uses a six-axis IMU (such as the LSM6DSO) integrating a three-axis accelerometer 21 and a three-axis gyroscope 22, with built-in motion detection interrupt functionality and a power consumption of less than 5μA in deep sleep mode. This module connects to the low-power microcontroller 1 via an SPI interface and connects its hardware interrupt pin to the wake-up interrupt input of the low-power microcontroller 1. The GNSS module 3 uses a high-precision module (such as the UM960) supporting dual-frequency differential positioning, achieving sub-meter positioning accuracy. It connects to the low-power microcontroller 1 via a UART interface and is powered by an independent power switch from the power management module 6. The communication module 4 uses a Cat1 module (such as the Air780E) supporting the three major telecom operators, connecting to the low-power microcontroller 1 via UART and a power enable pin. The water immersion sensor 5 is a contact electrode sensor, mounted on the bottom of the housing 8. Its pins are connected to the GPIO interrupt pin of the microcontroller 1, and it outputs an interrupt signal when conductive liquid is detected between the electrodes. The power management module 6 consists of multiple LDO power chips, each with an independent enable control pin (for electrical connection to the low-power microcontroller 1, inertial sensor module 2, GNSS module 3, communication module 4, and water immersion sensor 5), and its on / off state is controlled by the low-power microcontroller 1. The lithium-thionyl chloride battery 7 uses a 19Ah ER34615 type disposable lithium-thionyl chloride battery. The housing 8 is made of engineering plastic injection molding and has an IP67 or higher protection rating. It has an independent battery compartment inside. The inertial sensor module 2 is rigidly fixed to the central axis inside the housing 8 via a PCB to ensure that its measuring axis is consistent with the geometric axis of the cone.

[0026] The low-power microcontroller 1 is equipped with a device lifecycle state machine, which includes at least the following states: The deployed status S1 indicates that the cones have been stably placed in the work area; The S5 status during transport indicates that the cone has been confirmed to be retrieved and moved, and the cone continues to move. The S6 status indicates that the device has been stationary in a non-deployment position for an extended period of time, and its return to the warehouse has been confirmed. The S7 indicates that the device has entered deep sleep mode after a long period of inactivity.

[0027] The state machine reports the corresponding command word message during state transitions: when transitioning from the "moving" state S5 to the "reclaimed" state S6, it reports the command word "collected" message; when transitioning from the "reclaimed" state S6 to the "offline" state S7, it reports the command word "offline" message.

[0028] The 4G Cat1 communication module 4 can be replaced with NB-IoT, Cat-M1, LoRa, or satellite short message modules to adapt to different network coverage environments. The sub-meter level GNSS positioning module 3 can be replaced with a high-precision positioning module that supports multiple systems such as BeiDou RTK, Galileo, and GLONASS.

[0029] In addition to traffic cones, this intelligent cone action recognition device based on multi-level feature discrimination can also be applied to other temporary traffic facilities, such as construction warning signs, temporary guardrails, and anti-collision water barriers, as well as other IoT terminals with similar action recognition requirements. Example 2

[0030] This embodiment provides an intelligent cone action recognition method based on multi-level feature discrimination, such as... Figure 3 As shown, it includes the following steps: Step S1 (First-level feature discrimination – motion wake-up discrimination): After the device is powered on, it enters deep sleep mode, monitoring the acceleration signal of the cone only through the inertial sensor module 2 in low-power mode (continuously acquiring triaxial acceleration at a frequency of 25Hz). Specifically: the inertial sensor module 2 internally calculates the instantaneous change in resultant acceleration Δa and compares it with the first-level dynamic threshold Th1; when Δa > Th1 (the initial value of Th1 can be set to 0.3g), a hardware interrupt is triggered, waking up the low-power microcontroller 1 (that is, when the instantaneous change in resultant acceleration exceeds the first-level dynamic threshold Th1, a first-level wake-up is triggered, proceeding to step S2). Dynamic updates are then performed according to step S7.

[0031] Step S2 (Secondary Feature Discrimination – Data Acquisition and Feature Extraction): After the low-power microcontroller 1 is woken up, it immediately switches the inertial sensor module 2 to a high-frequency sampling mode (e.g., 208Hz, sampling interval Δt = 1 / fs ≈ 4.8ms), continuously acquiring triaxial acceleration data and triaxial gyroscope data (e.g., 208Hz, sampling interval Δt = 1 / fs ≈ 4.8ms) within a time window T (2 seconds). Figure 4 As shown), the basic signal is defined as follows: (a) Triaxial acceleration: φ(t) = (a x (t), a y (t), a z (t), unit: g; Where: g = 9.8 m / s 2 a x a y a z The three-axis acceleration components in the body coordinate system; (b) Triaxial angular velocity: ω(t) = (ωt) x (t), ω y (t), ω z(t), unit: ° / s; Where: ω x ω y ω z The three-axis angular velocity components in the body coordinate system; (c) Resultant acceleration modulus: |a(t)| = √(a x 2 (t) + a y 2 (t) + a z 2 (t) ); The magnitude of the resultant acceleration reflects the instantaneous magnitude of the acceleration vector; (d) Resultant angular velocity modulus: |ω(t)| = √(ω x 2 (t) + ω y 2 (t) + ω z 2 (t) ); The magnitude of the resultant angular velocity reflects the instantaneous magnitude of the angular velocity vector; (e) Instantaneous change in acceleration: Δa(t) = ||a(t)| - |a(t - Δt)||; The instantaneous change in acceleration is the fundamental quantity for triggering the first-level feature discrimination (step S1); (f) The attitude angle θ(t) is calculated in real time using a complementary filtering algorithm, and the formula is as follows: Attitude angle θ(t) = α · [θ(t-Δt) + ω x (t) · Δt ]+(1-α) · atan2( a y (t),√(a) x 2 (t) + a z 2 (t) ) ) Where: α is the filtering coefficient, and in this invention, α is taken as 0.98; θ(t) is in radians, and should be converted to degrees (°) as needed.

[0032] Within this time window, the following features are extracted (i.e., a multidimensional feature vector F, which includes time-domain features, frequency-domain features, attitude-domain features, and displacement features), as detailed below: (1) The peak value of the resultant acceleration A_peak = max(√(ax²+ay²+az²)); (2) Mean value of resultant acceleration ; (3) Peak-to-average ratio of combined acceleration (PMR) = A_peak / A_mean; (4) The main peak frequency F_main and the energy proportion E_h of the high-frequency band (>20Hz) after the resultant acceleration signal is transformed by FFT; F_main = ( fs / N ) · arg max_k P(k), k ∈ [ 1, N / 2 ]; E_h = ( Σ_{k : f_k>f_th} P(k) ) / ( Σ_{k=1}^{N / 2} P(k) ), where the cutoff frequency f_th = 20Hz is taken in this invention; the larger E_h is, the more prominent the high-frequency components of the signal are, corresponding to the "rigid impact" characteristic; (5) The cumulative change in attitude angle Δθ obtained by gyroscope integration: Δθ = |θ(t0+Tw) - θ(t0)|; (6) Peak rate of attitude angle change ω_peak: ω_peak = max{ |ω(t)| : t ∈ W}; (7) Duration of motion T_motion (i.e., the duration during which the resultant acceleration is greater than Th1 / 2): T_motion = ∫_W { Δa(t)>Th1 / 2} dt Where: {·} is an indicator function, which takes the value 1 when the condition is true and 0 otherwise; Th1 is the first-level dynamic threshold; (8) Attitude angle θ_end at the end of the window (0.5 seconds): θ_end = (1 / N_end) · Σ i θ(t i ), t i ∈ 0.5s from the end of the window; (9) Vibration attenuation coefficient Decay at the end of the window (0.5 seconds): Decay = - ln( A_RMS_end / A_peak ) / ( t_end - t_peak ); Where: t_peak is the peak time; A_RMS_end is the root mean square value of the resultant acceleration within 0.5s at the end of the window; it reflects the rate of vibration decay after impact. The faster the decay, the larger the Decay value, which corresponds to the action characteristic of "putting down and stabilizing". (10) Vibration stability σ_end: σ_end = √( (1 / N_end) · Σ i ( |a(t i )| - A_mean_end ) 2 ); Where: t iTake samples within 0.5s of the end of the window, where N_end is the number of samples in that sub-window; the smaller σ_end is, the more stable the vibration is at the end, which is the core basis for the "stable" determination; (11) Discrete Fourier Transform X(k): X(k) = Σ n =0 N-1 |a(t n )| · exp( -j · 2π · k · n / N ), k = 0, 1, …,N-1 Specifically, this invention is implemented on the MCU side using a radix-2 FFT algorithm, where N = 2. m (Typically, N = 256 or 512); (12) Power spectral density P(k): P(k) = |X(k)| 2 ; (13) Spectral entropy H_spec: H_spec = -Σ_k p(k) · log2p(k), p(k) = P(k) / Σ j P(j); It reflects the complexity of the spectrum; the larger the H_spec, the more uniform the frequency distribution (such as continuous oscillation), and the smaller the H_spec, the more concentrated the energy is at a certain frequency (such as rigid vibration). (14) Termination attitude stability σ_θ: σ_θ = √( (1 / N_end) · Σ i ( θ(t i ) - θ_end ) 2 ); (15) Attitude change path length L_θ: L_θ = Σ n =1 N-1 | θ(t n+1 ) - θ(t n ) |.

[0033] Step S3 (Action Classification and Determination): Input the multi-dimensional feature vector F into the preset multi-level discriminant model (e.g., ... Figure 6 As shown), calculate the confidence score for each type of action and output the action classification result; the action classification includes at least six categories: deployment and placement, retraction and removal, vehicle collision, illegal movement, natural dumping, and abnormal water immersion (e.g., Figure 5 (as shown) Specifically, it can be: (a) If A_peak>4g and Eh>60% and Δθ>30°, it is determined to be a "vehicle collision"; (b) If T_motion > 5 seconds and θ_end > 60° (still in the swinging state), it is determined to be the intermediate process of "retrieving and moving away"; (c) If T_motion < 3 seconds and θ_end < 15° (close to upright) and Decay < 0.1, it is determined as "deployed and put down"; (d) If T_motion < 5 seconds and the overall displacement is greater than the threshold and θ_end < 30°, it is judged as "illegal movement"; (e) If θ_end>70° (close to horizontal) and ω_peak<50° / s, it is judged as "natural tilting"; (f) If the water immersion sensor (5) outputs a valid signal, it is determined to be "abnormal water immersion"; This embodiment also includes a secondary confirmation window mechanism: When the confidence scores of all action classifications D_k(F) < τ_min, the system enters a pending confirmation state with a duration of T_conf = 30s. After reaching T_conf, the final state attitude is determined: if |θ_end|<30° and σ_θ<5°, it is determined as an illegal movement; If |θ_end|>60° or the movement continues, it is determined that the object should be retrieved and moved away. All other cases are classified as natural spillage.

[0034] Step S4 (Secondary wake-up decision): Determine whether location reporting and event reporting are required based on the action classification results. Only proceed to the next step if the determination result is that the event needs to be reported; otherwise, return to step S1 to continue monitoring. (1) If the cause is determined to be “deployment and placement”, “vehicle collision”, “illegal movement”, “natural dumping” or “abnormal water immersion”, then proceed to step S5; (2) If it is determined to be the intermediate process of "recovery and removal", then no report is made, and the inertial sensor module 2 is switched to the continuous monitoring mode with medium power consumption. After the movement stops, feature extraction is performed again, and the stable state is confirmed before deciding whether to report. (3) If it is determined to be a meaningless trigger such as environmental vibration, then proceed to step S7 to update the threshold and return to step S1.

[0035] Step S5 (Level 3 Wake-up - Positioning Acquisition): Start the GNSS module (i.e., the positioning module) to acquire the precise latitude and longitude coordinates of the cone; specifically, the low-power microcontroller 1 turns on the power of the sub-meter level GNSS positioning module 3 through the power management module 6, waits for it to complete positioning (typical time is 30-60 seconds), and immediately turns off its power after acquiring the latitude and longitude coordinates.

[0036] Step S6 (Level 4 Wake-up – Data Reporting): The communication module is activated, and the event type, location data, and device status information are packaged and uploaded to the cloud server. Specifically, the low-power microcontroller 1 powers on the 4G Cat1 communication module 4 via the power management module 6. After establishing a network connection, the following data is packaged and uploaded to the cloud server: unique device ID, event type code, latitude and longitude, timestamp, attitude angle, battery voltage, and signal strength. The 4G module power is immediately turned off after reporting is complete.

[0037] Step S7: Adaptively update the first-level dynamic threshold Th1 according to the following formula (e.g.) Figure 7 As shown): Th1=max(0.2g, Δa 95 ×1.5), where Δa 95 It represents the 95th percentile of the instantaneous change in the combined acceleration over the most recent hour.

[0038] In this embodiment, in step S2, the displacement feature includes at least the fused displacement Δd_hybrid, which is obtained by weighting the GNSS displacement and the displacement calculated by inertial dead reckoning. The formula for calculating the displacement characteristic Δd_hybrid is as follows: Δd_hybrid = K_g · Δd_gnss + (1-K_g) ·Δd_ins; K_g = HDOP_norm × Sats_norm; In the formula: K_g is the GNSS confidence weight, Δd_gnss is the displacement based on GNSS, Δd_ins is the displacement based on inertial dead reckoning, HDOP_norm=1 / max(HDOP,1), Sats_norm=min(Sats / 6,1).

[0039] In step S3, for the two types of actions, "illegal movement" and "removal", a three-dimensional boundary discrimination criterion including termination posture stability, displacement distance and motion continuity is used for fine classification.

[0040] The stability determination of the terminated attitude is based on the attitude angle θ(t) and its standard deviation σ_θ(t) within the observation window T_obs after the action stops. The displacement distance discrimination D2 is based on comparing the fused displacement Δd_hybrid with the preset maximum allowable displacement distance D_max; The motion continuity determination D3 is based on the number of motion phases N_phases and the longest single-phase motion duration T_phase_i during the motion process; The overall score is obtained through weighted fusion. When S_illegal ≥ preset threshold τ_min, it is judged as "illegal movement"; otherwise, it is judged as "reclaim and move" or enters the secondary confirmation window.

[0041] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent cone action recognition based on multi-level feature discrimination, characterized in that, Includes the following steps: Step S1: Monitor the acceleration signal of the cone in low-power mode using the inertial sensor module. When the instantaneous change in the resultant acceleration exceeds the first-level dynamic threshold Th1, trigger the first-level wake-up and proceed to step S2. Step S2: Activate the high-frequency sampling mode of the inertial sensor module, continuously collect acceleration data and gyroscope data within a preset time window, and extract the multi-dimensional feature vector F; The feature vector F includes time-domain features, frequency-domain features, attitude-domain features, and displacement features; Step S3: Input the multidimensional feature vector F into a preset multi-level discrimination model, calculate the confidence score of each type of action, and output the action classification result; the action classification includes at least six categories: deployment and placement, retraction and removal, vehicle collision, illegal movement, natural dumping and abnormal water immersion. Step S4: Determine whether location reporting and event reporting are required based on the action classification results. Only proceed to the next step if the determination result is that the event needs to be reported; otherwise, return to step S1 to continue monitoring. Step S5: Activate the GNSS module to obtain the precise latitude and longitude coordinates of the cone; Step S6: Start the communication module and upload the event type, location data, and device status information to the cloud server. Step S7: Adaptively update the first-level dynamic threshold Th1 according to the following formula: Th1=max(0.2g, Δa) 95 ×1.5), where Δa 95 It represents the 95th percentile of the instantaneous change in the combined acceleration over the most recent hour.

2. The intelligent cone action recognition method based on multi-level feature discrimination according to claim 1, characterized in that: In step S2, the time-domain features include the peak resultant acceleration A_peak, the mean resultant acceleration A_mean, the peak-to-average power ratio (PMR) of the resultant acceleration, the vibration attenuation coefficient Decay, the vibration stability σ_end, and the motion duration T_motion; the frequency-domain features include the discrete Fourier transform X(k), the power spectral density P(k), the main peak frequency F_main, the high-frequency energy proportion E_h, and the spectral entropy H_spec; the attitude-domain features include the cumulative change in attitude angle Δθ, the peak rate of attitude angle change ω_peak, the final attitude angle θ_end, the final attitude stability σ_θ, and the attitude change path length L_θ.

3. The intelligent cone action recognition method based on multi-level feature discrimination according to claim 1 or 2, characterized in that: In step S2, the displacement feature includes at least the fused displacement Δd_hybrid, which is obtained by weighting the GNSS displacement and the displacement calculated by inertial dead reckoning. The formula for calculating the displacement characteristic Δd_hybrid is as follows: Δd_hybrid = K_g · Δd_gnss + (1-K_g) ·Δd_ins; K_g = HDOP_norm × Sats_norm; In the formula: K_g is the GNSS confidence weight, Δd_gnss is the displacement based on GNSS, Δd_ins is the displacement based on inertial dead reckoning, HDOP_norm=1 / max(HDOP,1), Sats_norm=min(Sats / 6,1).

4. The intelligent cone action recognition method based on multi-level feature discrimination according to claim 1 or 2, characterized in that: In step S3, for the two types of actions, "illegal movement" and "removal", a three-dimensional boundary discrimination criterion including termination posture stability, displacement distance and motion continuity is used for fine classification.

5. The intelligent cone action recognition method based on multi-level feature discrimination according to claim 4, characterized in that: The stability determination of the terminated attitude is based on the attitude angle θ(t) and its standard deviation σ_θ(t) within the observation window T_obs after the action stops. The displacement distance discrimination D2 is based on comparing the fused displacement Δd_hybrid with the preset maximum allowable displacement distance D_max; The motion continuity determination D3 is based on the number of motion phases N_phases and the longest single-phase motion duration T_phase_i during the motion process; The comprehensive score S_illegal is obtained by weighted fusion, which is calculated as w1·D1 + w2·D2 + w3·D3. When S_illegal is greater than or equal to the preset threshold τ_min, it is judged as "illegal movement". Otherwise, it is judged as "reclaim and move" or enters the secondary confirmation window.

6. The intelligent cone action recognition method based on multi-level feature discrimination according to claim 5, characterized in that, Step S3 also includes a secondary confirmation window mechanism: When the confidence scores of all action classifications D_k(F) < τ_min, the system enters a pending confirmation state with a duration of T_conf=30s. Upon reaching T_conf, the final state attitude is determined: if |θ_end| < 30° and σ_θ < 5°, it is considered an illegal movement; If |θ_end| > 60° or the movement continues, it is determined that the object should be retrieved and moved away. All other cases are classified as natural spillage.

7. A smart cone action recognition device based on multi-level feature discrimination, characterized in that, include: Low-power microcontroller (1); An inertial sensor module (2) is electrically connected to the low-power microcontroller (1) and is used to collect the acceleration and angular velocity data of the cone. GNSS module (3), which is electrically connected to the low-power microcontroller (1), is used to obtain the latitude and longitude coordinates of the cone; The communication module (4) is electrically connected to the low-power microcontroller (1) and is used to upload data to the cloud server; A water immersion sensor (5) is electrically connected to a low-power microcontroller (1); The power management module (6) is electrically connected to the low-power microcontroller (1), the inertial sensor module (2), the GNSS module (3), the communication module (4), and the water immersion sensor (5); A lithium-ion battery (7) is electrically connected to the power management module (6); Housing (8) for accommodating the above-mentioned components; the water immersion sensor (5) is disposed at the bottom of housing (8); The low-power microcontroller (1) is configured to perform the intelligent cone action recognition method according to any one of claims 1 to 6.

8. The intelligent cone action recognition device based on multi-level feature discrimination according to claim 8, characterized in that: The low-power microcontroller (1) is equipped with a device lifecycle state machine, which includes at least the following states: The deployed status S1 indicates that the cones have been stably placed in the work area; The S5 status during transport indicates that the cone has been confirmed to be retrieved and moved, and the cone continues to move. The S6 status indicates that the device has been stationary in a non-deployment position for an extended period of time, and its return to the warehouse has been confirmed. The S7 indicates that the device has entered deep sleep mode after a long period of inactivity.

9. The intelligent cone action recognition device based on multi-level feature discrimination according to claim 8, characterized in that: The state machine reports the corresponding command message during state transitions: When the process transitions from the moving state S5 to the recovered state S6, a command word "collected" message is reported. When transitioning from the recovered state S6 to the offline state S7, an offline command message is reported.

10. The intelligent cone action recognition device based on multi-level feature discrimination according to claim 8, characterized in that, The inertial sensor module (2) includes an accelerometer (21) and a gyroscope (22), which are rigidly fixed on the central axis inside the housing (8) to ensure that its measuring axis is consistent with the geometric axis of the cone; the protection level of the housing (8) reaches IP67 or above.

Images