Intelligent detection method and system for bar collision of car stop based on deep learning fusion

Abstract

The invention discloses a deep learning fusion-based intelligent detection method and system for bar collision of a car stop, and belongs to the technical field of safety monitoring of intelligent traffic facilities. The method comprises the steps of collecting an original time sequence data stream of a car stop rod in real time, carrying out preprocessing and feature extraction on the original time sequence data stream, constructing a multi-dimensional feature time sequence, inputting the multi-dimensional feature time sequence into a pre-trained deep learning behavior recognition model, calculating a classification probability and a feature embedding vector, and carrying out feature extraction on the classification probability and the feature embedding vector. And performing real-time similarity comparison on the abnormal confidence coefficient and a template in a reference behavior feature library, generating an abnormal confidence coefficient score according to a self-adaptive dynamic threshold value strategy, and if the abnormal confidence coefficient score continuously exceeds a self-adaptive dynamic threshold value and accords with a preset collision behavior time sequence mode, judging that a rod collision behavior occurs, and starting a linkage evidence obtaining and alarm process. The problems that a traditional scheme is high in false alarm rate and poor in environmental adaptability are solved, and accurate and intelligent detection and automatic evidence obtaining of the rod collision behavior are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation facility safety monitoring technology, specifically to an intelligent detection method and system for vehicle barrier collision barriers based on deep learning fusion. Background Technology

[0002] With the significant increase in the number of motor vehicles and the development of intelligent parking lot entrances and exits, parking lot entrance and exit systems are gradually becoming unmanned. However, during this process, instances of people forcibly raising or vehicles hitting the barriers have occurred, resulting in damage to the parking lane barriers. With no one in the lane, parking lot management cannot be notified in a timely manner, and vehicles may leave without paying. Furthermore, damage to the barriers cannot be traced for compensation, causing economic losses to parking lot system owners. To prevent these situations, parking lot entrance and exit system equipment needs to detect instances of people forcibly raising or vehicles hitting the barriers and link this to lane environment monitoring for evidence collection and processing.

[0003] Currently, the detection of collision barriers at parking lot entrances and exits mainly relies on manual inspections of video surveillance or simple logical judgments by vehicle detectors, which has the following drawbacks: 1. Poor real-time performance, unable to prevent timely intervention; 2. High false alarm rate, easily affected by light and weather conditions; 3. It is impossible to accurately distinguish between intentional impacts and normal contact or environmental interference. Some solutions also use vibration sensors, but they are difficult to identify the source and force of the impact, making it impossible to effectively collect evidence.

[0004] To address the aforementioned issues, there is an urgent need for an intelligent detection method and system for vehicle barrier collision barriers based on deep learning fusion, which can solve the problems existing in traditional methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for intelligent detection of vehicle barrier collisions based on deep learning fusion, which solves the problems of high false alarm rate and poor environmental adaptability of traditional solutions, and realizes accurate and intelligent detection and automatic evidence collection of barrier collision behavior.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for intelligent detection of collision barriers based on deep learning fusion includes: Step 1: Real-time acquisition of the raw time-series data stream of the barrier arm; Step 2: Preprocess and extract features from the original time-series data stream to construct a multi-dimensional feature time-series sequence representing the motion state of the barrier arm; Step 3: Input the multi-dimensional feature time series into the pre-trained deep learning behavior recognition model to calculate the classification probability and feature embedding vector of the current barrier lever's operating state. The deep learning behavior recognition model is built based on an improved 1D-CNN network structure and an LSTM network structure. The specific improvements are as follows: multi-scale convolutional kernels are used on the standard 1D-CNN network structure to extract local features under different receptive fields; skip connections are added on the standard 1D-CNN network structure to fuse shallow and deep features; an attention mechanism is introduced on the standard LSTM network structure; and temporal pyramid pooling is set after the output layer of the LSTM network structure with the attention mechanism to achieve adaptive pooling for temporal inputs of different lengths. Step 4: Perform real-time similarity comparison between the feature embedding vector at the current moment and the template in the benchmark behavior feature library, and generate anomaly confidence scores based on the adaptive dynamic threshold strategy; Step 5: If the abnormal confidence score continues to exceed the adaptive dynamic threshold and meets the preset impact behavior timing pattern, it is determined that a pole collision has occurred, and the linkage evidence collection and alarm process is initiated.

[0007] Furthermore, in step 1, the raw time-series data stream of the barrier arm is collected in real time, specifically as follows: A multi-dimensional sensor module is installed on the barrier arm, and the original time-series data stream of the barrier arm is collected based on the multi-dimensional sensor module. The multi-dimensional sensor module includes at least an IMU (Inertial Measurement Unit).

[0008] Furthermore, in step 2, the original time-series data stream is preprocessed and features are extracted to construct a multi-dimensional feature time-series sequence representing the motion state of the barrier arm, specifically as follows: Based on the Kalman filter and complementary filter algorithm, the original time series data stream is fused and denoised to obtain stable pole attitude angle time series data; Temporal statistical features, frequency statistical features, and trajectory features based on sliding windows are extracted from the time-series data of the pole's attitude angles. Multidimensional feature time series sequences are constructed based on time-domain statistical features, frequency-domain statistical features, and trajectory features based on sliding windows.

[0009] Furthermore, in step 4, the feature embedding vector at the current moment is compared with the template in the benchmark behavior feature library in real time, and an anomaly confidence score is generated according to the adaptive dynamic threshold strategy, specifically as follows: Constructing and maintaining a benchmark behavior feature library: Collecting operational data of the barrier arm under various standard working conditions, extracting corresponding feature embedding vectors through a deep learning behavior recognition model, and storing them as a feature template set according to working conditions; Perform real-time similarity comparison: calculate the similarity between the feature embedding vector at the current moment and each feature template set in the benchmark behavior feature library, and take the highest similarity value as the matching degree between the current motion mode and the known normal or known interference mode; Calculate the adaptive dynamic threshold: Set a basic judgment threshold and calculate a dynamic adjustment factor based on real-time environmental sensing data; multiply the basic judgment threshold and the dynamic adjustment factor to obtain the adaptive dynamic threshold at the current time. Generate anomaly confidence score: Subtract the value 1 from the highest similarity value, and the difference is used as the anomaly confidence score; compare the anomaly confidence score with the adaptive dynamic threshold, and if it exceeds the threshold, the current moment is determined to be an instantaneous anomaly.

[0010] Furthermore, in step 5, the initiation of the linkage evidence collection and alarm process specifically includes: Triggering high-definition surveillance cameras to capture high-resolution images and video clips of the event scene; Automatically extract license plate information from captured media using image recognition algorithms; The event log is stored in the database, and an alarm message is sent to the upper-level management system.

[0011] This invention also provides a deep learning-based intelligent detection system for wheel chock collisions, applied to the aforementioned deep learning-based intelligent detection method for wheel chock collisions, comprising: The sensor data acquisition module is fixedly installed on the barrier arm and is used to acquire the raw time-series data stream of the barrier arm in real time. The sensor data acquisition module includes at least an IMU inertial measurement unit. The edge computing and intelligent analysis module, which is connected to the sensor data acquisition module, is used to receive the raw time-series data stream and realize the detection of the barrier arm. The linkage control and evidence collection module, which is connected to the edge computing and intelligent analysis module, is used to initiate the high-definition camera capture, license plate recognition and alarm process after receiving the collision behavior judgment signal; The data communication module connects to the edge computing and intelligent analysis module and the linkage control and evidence collection module, respectively, and is used to interact with the cloud server or parking management platform.

[0012] Furthermore, the edge computing and intelligent analysis module specifically includes: The data preprocessing unit is used to filter, fuse, and extract features from the original time-series data stream to construct a multi-dimensional feature time-series sequence. The model inference unit contains a pre-trained deep learning behavior recognition model, which is used to process multi-dimensional feature time series and output classification probability and feature embedding vector. The intelligent decision-making unit has a built-in benchmark behavioral feature library, which is used to perform real-time similarity comparison of feature vectors, adaptive dynamic threshold calculation, and abnormal confidence score generation.

[0013] Furthermore, the linkage control and evidence collection module specifically includes: The camera triggering unit is used to send a trigger command to a preset high-definition network camera after receiving a collision behavior determination signal; The license plate recognition unit is used to process media files captured by the camera and automatically extract license plate information from them; The alarm and recording unit is used to generate structured event records and send alarm messages, while controlling on-site audible and visual alarm devices.

[0014] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention significantly improves the accuracy of pole collision behavior recognition and greatly reduces the false alarm rate: By integrating multi-dimensional sensors and an improved deep learning model, the system can extract discriminative deep features from complex temporal motion data, achieving accurate differentiation between normal passage, environmental interference, and malicious collision behavior. Compared with traditional detection methods, this method exhibits extremely high detection accuracy and stability in real-world complex scenarios, fundamentally suppressing false alarms.

[0015] 2. This invention possesses superior environmental adaptability and anti-interference capabilities: The innovatively introduced adaptive dynamic threshold mechanism intelligently adjusts the judgment strategy based on real-time environmental conditions, effectively distinguishing between pole disturbances caused by environmental factors such as wind and actual impact events. This system maintains reliable operation under various weather and operating conditions, overcoming the common problem of frequent false triggering in traditional solutions in variable environments.

[0016] 3. This invention achieves fully automated evidence collection and rapid response: The system has a built-in intelligent linkage mechanism. Once the collision is confirmed, it can automatically trigger high-definition image capture, license plate recognition, evidence preservation, and alarm notification, forming a complete automated event handling closed loop. This process requires no manual intervention and can complete the entire process from perception and decision-making to handling in a very short time, greatly improving security management efficiency and event traceability capabilities. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0019] like Figure 1 As shown, this invention provides an intelligent detection method for vehicle barrier collisions based on deep learning fusion, comprising: Step 1: Real-time acquisition of the raw time-series data stream of the barrier arm; Step 2: Preprocess and extract features from the original time-series data stream to construct a multi-dimensional feature time-series sequence representing the motion state of the barrier arm; Step 3: Input the multidimensional feature time series into the pre-trained deep learning behavior recognition model to calculate the classification probability and feature embedding vector of the current barrier lever's operating state; Step 4: Perform real-time similarity comparison between the feature embedding vector at the current moment and the template in the benchmark behavior feature library, and generate anomaly confidence scores based on the adaptive dynamic threshold strategy; Step 5: If the abnormal confidence score continues to exceed the adaptive dynamic threshold and meets the preset impact behavior timing pattern, it is determined that a pole collision has occurred, and the linkage evidence collection and alarm process is initiated.

[0020] In step 1, the raw time-series data stream of the barrier arm is collected in real time, specifically as follows: A multi-dimensional sensor module is installed on the barrier arm, and the original time-series data stream of the barrier arm is collected based on the multi-dimensional sensor module. The multi-dimensional sensor module includes at least an IMU (Inertial Measurement Unit).

[0021] In step 2, the original time-series data stream is preprocessed and features are extracted to construct a multi-dimensional feature time-series sequence representing the motion state of the barrier arm, specifically as follows: Step 201: The original time-series data stream is fused and denoised based on Kalman filtering and complementary filtering algorithms to obtain stable pole attitude angle time-series data. The Kalman filtering and complementary filtering algorithms will be described in detail below: 1. Kalman Filter Kalman filtering is a recursive state estimation algorithm suitable for linear Gaussian systems. It makes the optimal estimate of the system state through two steps: prediction and update. State equations and observation equations: (1) State vector (taking attitude angle as an example): (1) Where θ is the attitude angle (pitch / roll). Angular velocity; (2) State prediction equation: (2) (3) in, F t Here is the state transition matrix. P Let be the error covariance matrix. Q t For process noise covariance, The control input matrix represents the degree of influence of external control variables on the system state. In this invention, the control input vector represents the equivalent control quantity of the driving torque of the barrier motor on the pole.

[0022] (3) Observation update equation: (4) (5) (6) in, H t For the observation matrix, z t For observations (such as the gravitational component measured by an accelerometer). R t To observe the noise covariance, K t For Kalman gain.

[0023] 2. Complementary Filter Complementary filtering achieves a smooth attitude angle by fusing high-frequency gyroscope data with low-frequency stable accelerometer data.

[0024] The basic formula is: (7) in, θ t The attitude angle at the current moment, ω t The angular velocity of the gyroscope. θ acc The attitude angles calculated by the accelerometer (such as through...) ), α The filter coefficients (usually around 0.98) control the weights of the two, Δ t The sampling interval is denoted as .

[0025] Step 202: Extract time-domain statistical features, frequency-domain statistical features, and sliding window-based trajectory features from the time-series data of the pole's attitude angles. These features are described below: 1. Time-domain statistical characteristics (1) Mean (8) (2) Variance (9) (3) Extreme values (10) 2. Frequency domain statistical characteristics The fast Fourier transform (FFT) of the attitude angle timing signal x(n) is as follows: (11) Extracting the main frequency component: Find the frequency corresponding to the maximum value in the amplitude spectrum |X(k)|. ; The first m main frequency components and their amplitudes can be extracted.

[0026] 3. Trajectory features based on sliding window (1) Sliding window mean (12) (2) Sliding window standard deviation (13) (3) Trajectory curvature (approximate calculation) (14) In the formula, .

[0027] Step 203: Construct a multidimensional feature time series based on time-domain statistical features, frequency-domain statistical features, and trajectory features based on a sliding window; In summary, the above steps can be summarized as follows: 1. Data Acquisition: IMU outputs raw acceleration. a and angular velocity ω .

[0028] 2. Filtering and fusion: Use Kalman filtering ω Perform state estimation; Complementary filtering is used to fuse accelerometer and gyroscope data to obtain smooth attitude angles. θ .

[0029] 3. Feature extraction: from θ Extracting temporal statistical features from sequences; right θ Perform FFT to extract frequency domain features; Calculate trajectory features (such as curvature and trend) within a sliding window.

[0030] 4. Output: Constructs a multi-dimensional feature time sequence, which is then input into the deep learning behavior recognition model.

[0031] In step 3, the multidimensional feature time series is input into a pre-trained deep learning behavior recognition model to calculate the classification probability and feature embedding vector of the current barrier lever's operating state, specifically: A deep learning behavior recognition model is constructed based on the improved 1D-CNN network structure and LSTM network structure; The deep learning behavior recognition model is trained based on a pre-set dataset; By inputting the multidimensional feature time series into the trained deep learning behavior recognition model, the classification probability and feature embedding vector of the current barrier lever's operating state are obtained, along with the probability distribution of multiple categories, including: Category 1: Normal landing gate; Category 2: Wind-induced vibration; Category 3: Minor impact; Category 4: Severe impact.

[0032] The improved 1D-CNN and LSTM network structures are as follows: Based on the standard 1D-CNN network structure, multi-scale convolutional kernels (such as parallel lengths of 3, 5, and 7) are used to extract local features under different receptive fields. Skip connections are added to the standard 1D-CNN network structure to fuse shallow and deep features, alleviate gradient vanishing, and improve feature reuse capability. An attention mechanism is introduced into the standard LSTM network structure, enabling the model to focus on features at key time steps and improve its sensitivity to collision behavior. A specific formula example is given below: (15) in, h t For LSTM hidden states, h s For context vectors; Temporal pyramid pooling (TPP) is set after the output layer of the LSTM network structure with an attention mechanism to achieve adaptive pooling for temporal inputs of different lengths.

[0033] The standard 1D-CNN and standard LSTM network structures are existing technologies and will not be described in detail here. The following will focus on explaining the training process: 1. Data preparation and annotation A large amount of data on the operation of the vehicle stop devices was collected and labeled into four states by manual or semi-automatic methods.

[0034] Dataset split: 70% training set, 15% validation set, and 15% test set.

[0035] 2. Loss Function Using cross-entropy loss: (16) in For real labels, To predict probabilities.

[0036] 3. Optimizer and Hyperparameters Optimizer: Adam, initial learning rate lr=0.001.

[0037] Batch size: 32.

[0038] Training rounds: 100, early stopping strategy (stop when the validation set loss does not decrease for 5 consecutive rounds).

[0039] 4. Model Evaluation Indicators Accuracy, precision, recall, and F1 score.

[0040] Confusion matrix analysis focuses on the recognition performance of "collision" type.

[0041] In step 4, the feature embedding vector at the current moment is compared with the template in the benchmark behavior feature library in real time, and an anomaly confidence score is generated according to the adaptive dynamic threshold strategy, specifically: Step 401: Constructing and Maintaining a Benchmark Behavior Feature Library: Collect operational data of the barrier arm under various standard working conditions, extract corresponding feature embedding vectors through a deep learning behavior recognition model, and store them as a feature template set according to working conditions, specifically: Data Acquisition: During the initial deployment or calibration of the system, the barrier is actively operated under controlled conditions. Time-series data covering multiple "standard operating conditions" are collected, primarily including: (a) multiple normal automatic barrier raising and lowering processes; and (b) data on barrier stationary and slight swaying under different wind speeds (simulated by fans). This data constitutes the system's benchmark for "normal" and "known non-malicious interference".

[0042] Feature extraction and storage: The collected time-series data, after preprocessing and feature extraction, is input into the trained deep learning behavior recognition model to extract the feature embedding vectors output by the model's temporal pyramid pooling layer.

[0043] Library Structure Organization: The extracted feature embedding vectors are categorized and stored according to their corresponding operating conditions, forming a structured feature template set. For example, multiple sets can be established, such as Template Set_Normal Pole Raising, Template Set_Normal Pole Lowering, and Template Set_Wind Force Level 1. Each set contains multiple feature vectors obtained from multiple samplings under that operating condition to characterize the normal fluctuation range of this type of behavior.

[0044] Step 402: Perform real-time similarity comparison: Calculate the similarity between the feature embedding vector at the current moment and each feature template set in the benchmark behavior feature library, and take the highest similarity value as the matching degree between the current motion mode and the known normal or known interference mode. The similarity calculation standard of this invention adopts pre-similarity. Step 403: Calculate the adaptive dynamic threshold: Set a basic judgment threshold and calculate a dynamic adjustment factor based on real-time environmental sensing data. The dynamic adjustment factor is adjusted based on environmental parameters. For example, set up environmental sensors to collect environmental data. In this invention, wind speed is used as an example, and real-time wind speed is collected by an anemometer. v Define a wind speed adjustment factor function as follows: (17) In the formula, v calm The calm wind speed threshold (e.g., 2 m / s). k This is the adjustment coefficient; The greater the wind speed, the higher the α wind The larger (t) is, the higher the final threshold becomes, and the more "lenient" the system's judgment becomes, thus avoiding false alarms caused by strong winds; Multiply the basic judgment threshold by the dynamic adjustment factor to obtain the adaptive dynamic threshold at the current time. Step 404: Generate anomaly confidence score: Subtract the value 1 from the highest similarity value, and use the difference as the anomaly confidence score; compare the anomaly confidence score with the adaptive dynamic threshold. If it exceeds the threshold, the current moment is determined to be an instantaneous anomaly.

[0045] Step 5, the initiation of the linkage evidence collection and alarm process, specifically includes: Step 501: Trigger the high-definition surveillance camera to capture high-resolution images and video clips of the event scene, specifically: Trigger mechanism: The embedded edge computing module sends a trigger signal to the high-definition network camera through GPIO (General-Purpose Input / Output) pins or a preset communication protocol (such as RS485, Ethernet). This signal is a low-level pulse or a specific message containing an event ID.

[0046] Snapshots strategy: The camera preset position is aligned with the area of the barrier gate rod. After being triggered, the camera performs the following: Capture 3 - 5 high-resolution (such as 1920×1080) JPEG pictures at an interval of 0.5 seconds.

[0047] Simultaneously record a 10 - 15-second H.264 encoded video clip (such as 5 seconds before the trigger and 10 seconds after the trigger) around the event.

[0048] All media files are marked with timestamps, location codes, and the unique event ID as the file name.

[0049] Step 502: Automatically extract the license plate information from the captured media through an image recognition algorithm. Specifically: Vehicle detection: Use the lightweight YOLOv5s model to locate the vehicle area in the captured pictures.

[0050] License plate location: Within the vehicle area, use a traditional image processing algorithm based on morphological operations and color segmentation or a small convolutional neural network to accurately locate the license plate area.

[0051] Character recognition: Send the license plate area image into a pre-trained license plate character recognition model. This model is based on the CRNN (Convolutional Recurrent Neural Network) structure to perform sequence recognition on the segmented characters.

[0052] Result output: Output a formatted string, such as "Beijing A·12345".

[0053] The license plate information detection method is relatively mature in the prior art. Therefore, only one embodiment is given in this invention as long as it can be implemented.

[0054] Step 503: Store the event record in the database and send an alarm message to the superior management system. Specifically: This invention provides a schematic table for the design of the database table structure, as shown in Table 1 specifically: Table 1 Database table structure design table This invention also provides a format of the alarm message (JSON example), which is: json { "event_id": "COL_20231027_142035_001", "timestamp": "2023-10-27 14:20:35", "location": "PARK_A_ENTRY_01", "alert_level": "HIGH", "behavior": "SEVERE_IMPACT", "confidence": 0.96, "media_urls": ["... / img1.jpg", "... / video.mp4"], "plate_number": "Beijing A·12345", "action_required": "IMMEDIATE_REVIEW" }.

[0055] The present invention also provides an intelligent detection system for the barrier rod of a parking barrier based on deep learning fusion, which is applied to the above-mentioned intelligent detection method for the barrier rod of a parking barrier based on deep learning fusion, and includes: A sensor data acquisition module, fixedly installed on the barrier rod of the parking barrier, is used to collect the original time-series data stream of the barrier rod in real time. The sensor data acquisition module at least includes an IMU inertial measurement unit; An edge computing and intelligent analysis module, communicatively connected to the sensor data acquisition module, is used to receive the original time-series data stream and implement the detection of the barrier rod collision of the parking barrier; A linkage control and evidence collection module, communicatively connected to the edge computing and intelligent analysis module, is used to start the processes of high-definition camera抓拍, license plate recognition and alarm after receiving the signal for determining the collision behavior; A data communication module, respectively connected to the edge computing and intelligent analysis module and the linkage control and evidence collection module, is used to perform data interaction with a cloud server or a parking lot management platform.

[0056] The edge computing and intelligent analysis module specifically includes: A data preprocessing unit, used to perform filtering fusion and feature extraction on the original time-series data stream, and construct a multi-dimensional feature time-series sequence; A model inference unit, embedded with a pre-trained deep learning behavior recognition model, is used to process the multi-dimensional feature time-series sequence and output a classification probability and a feature embedding vector; An intelligent decision-making unit, built with a benchmark behavior feature library, is used to perform real-time similarity comparison of feature vectors, calculate an adaptive dynamic threshold, and generate an abnormal confidence score.

[0057] The linkage control and evidence collection module specifically includes: The camera triggering unit is used to send a trigger command to a preset high-definition network camera after receiving a collision behavior determination signal; The license plate recognition unit is used to process media files captured by the camera and automatically extract license plate information from them; The alarm and recording unit is used to generate structured event records and send alarm messages, while controlling on-site audible and visual alarm devices.

[0058] The present invention also provides an embodiment, specifically a comparative test between the system of the present invention and a traditional vibration sensor solution, as follows: 1. Test environment: Location: Exit lane of a commercial parking lot.

[0059] Duration: 30 consecutive days.

[0060] Interference includes various scenarios such as strong winds, rain, normal vehicle traffic, and human-caused collisions.

[0061] Comparison system: Traditional alarm system based on a single vibration sensor (threshold trigger).

[0062] 2. Test Data and Results: Total number of triggering events: 12 actual collisions, 25 instances of strong wind interference, and over 3,000 instances of normal gate raising.

[0063] System performance of this invention: Impact behavior recognition accuracy: 12 / 12 = 100%; Number of false alarms for impact behavior (misjudging wind disturbance / normal behavior as impact): 1 time; False alarm rate: 1 / (3000+25) ≈ 0.03%; Average response time (from impact to alarm): <2 seconds; Accuracy rate of license plate information extraction: 11 / 12≈91.7%; Performance of traditional vibration sensor systems: Impact behavior recognition accuracy: 10 / 12≈83.3% (two minor impacts did not trigger the reaction); Number of false alarms for collisions: 89 (mainly caused by strong winds and heavy vehicles passing by); False alarm rate: 89 / (3000+25)≈2.94%; Average response time: ≈1 second (sensor-triggered only, relatively fast); License plate information extraction requires manual review of the video afterwards.

[0064] The comparative data above demonstrates that the method and system described in this invention significantly outperform traditional single vibration sensor solutions in terms of accuracy in impact behavior detection, extremely low false alarm rate, strong environmental anti-interference capability, and automated evidence collection. Particularly during the testing period with frequent strong winds, the system of this invention only generated one false alarm, while the traditional system generated as many as 89 false alarms. This effectively solves the core problems of high false alarm rate and poor environmental adaptability in existing technologies, achieving highly reliable safety protection.

[0065] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for intelligent detection of collision barriers based on deep learning fusion, characterized in that, include: Step 1: Real-time acquisition of the raw time-series data stream of the barrier arm; Step 2: Preprocess and extract features from the original time-series data stream to construct a multi-dimensional feature time-series sequence representing the motion state of the barrier arm; Step 3: Input the multi-dimensional feature time series into the pre-trained deep learning behavior recognition model to calculate the classification probability and feature embedding vector of the current barrier lever's operating state. The deep learning behavior recognition model is built based on an improved 1D-CNN network structure and an LSTM network structure. The specific improvements are as follows: multi-scale convolutional kernels are used on the standard 1D-CNN network structure to extract local features under different receptive fields; skip connections are added on the standard 1D-CNN network structure to fuse shallow and deep features; an attention mechanism is introduced on the standard LSTM network structure; and temporal pyramid pooling is set after the output layer of the LSTM network structure with the attention mechanism to achieve adaptive pooling for temporal inputs of different lengths. Step 4: Perform real-time similarity comparison between the feature embedding vector at the current moment and the template in the benchmark behavior feature library, and generate anomaly confidence scores based on the adaptive dynamic threshold strategy; Step 5: If the abnormal confidence score continues to exceed the adaptive dynamic threshold and meets the preset impact behavior timing pattern, it is determined that a pole collision has occurred, and the linkage evidence collection and alarm process is initiated.

2. The intelligent detection method for vehicle barrier collision bars based on deep learning fusion according to claim 1, characterized in that, In step 1, the raw time-series data stream of the barrier arm is collected in real time, specifically as follows: A multi-dimensional sensor module is installed on the barrier arm, and the original time-series data stream of the barrier arm is collected based on the multi-dimensional sensor module. The multi-dimensional sensor module includes at least an IMU (Inertial Measurement Unit).

3. The intelligent detection method for vehicle barrier collision bars based on deep learning fusion according to claim 2, characterized in that, In step 2, the original time-series data stream is preprocessed and features are extracted to construct a multi-dimensional feature time-series sequence representing the motion state of the barrier arm, specifically as follows: Based on the Kalman filter and complementary filter algorithm, the original time series data stream is fused and denoised to obtain stable pole attitude angle time series data; Temporal statistical features, frequency statistical features, and trajectory features based on sliding windows are extracted from the time-series data of the pole's attitude angles. Multidimensional feature time series sequences are constructed based on time-domain statistical features, frequency-domain statistical features, and trajectory features based on sliding windows.

4. The intelligent detection method for vehicle barrier collision bars based on deep learning fusion according to claim 3, characterized in that, In step 4, the feature embedding vector at the current moment is compared with the template in the benchmark behavior feature library in real time, and an anomaly confidence score is generated according to the adaptive dynamic threshold strategy, specifically: Constructing and maintaining a benchmark behavior feature library: Collecting operational data of the wheel stop lever under various standard working conditions, extracting corresponding feature embedding vectors through a deep learning behavior recognition model, and storing them as a feature template set according to working conditions; Perform real-time similarity comparison: calculate the similarity between the feature embedding vector at the current moment and each feature template set in the benchmark behavior feature library, and take the highest similarity value as the matching degree between the current motion mode and the known normal or known interference mode; Calculate the adaptive dynamic threshold: Set a basic judgment threshold and calculate a dynamic adjustment factor based on real-time environmental sensing data; multiply the basic judgment threshold and the dynamic adjustment factor to obtain the adaptive dynamic threshold at the current time. Generate anomaly confidence score: Subtract the value 1 from the highest similarity value, and the difference is taken as the anomaly confidence score; compare the anomaly confidence score with the adaptive dynamic threshold, and if it exceeds the threshold, the current moment is determined to be an instantaneous anomaly.

5. The intelligent detection method for vehicle barrier collision bars based on deep learning fusion according to claim 4, characterized in that, Step 5, the initiation of the linkage evidence collection and alarm process, specifically includes: Triggering high-definition surveillance cameras to capture high-resolution images and video clips of the event scene; Automatically extract license plate information from captured media using image recognition algorithms; The event log is stored in the database, and an alarm message is sent to the upper-level management system.

6. A deep learning-based intelligent detection system for wheel chock collisions, applied to the deep learning-based intelligent detection method for wheel chock collisions as described in any one of claims 1-5, characterized in that, include: The sensor data acquisition module is fixedly installed on the barrier arm and is used to acquire the raw time-series data stream of the barrier arm in real time. The sensor data acquisition module includes at least an IMU inertial measurement unit. The edge computing and intelligent analysis module, which is connected to the sensor data acquisition module, is used to receive the raw time-series data stream and realize the detection of the barrier arm. The linkage control and evidence collection module, which is connected to the edge computing and intelligent analysis module, is used to initiate the high-definition camera capture, license plate recognition and alarm process after receiving the collision behavior judgment signal; The data communication module connects to the edge computing and intelligent analysis module and the linkage control and evidence collection module, respectively, and is used to interact with the cloud server or parking management platform.

7. The intelligent detection system for vehicle barrier collision bars based on deep learning fusion according to claim 6, characterized in that, The edge computing and intelligent analysis module specifically includes: The data preprocessing unit is used to filter, fuse, and extract features from the original time-series data stream to construct a multi-dimensional feature time-series sequence. The model inference unit contains a pre-trained deep learning behavior recognition model, which is used to process multi-dimensional feature time series and output classification probability and feature embedding vector. The intelligent decision-making unit has a built-in benchmark behavioral feature library, which is used to perform real-time similarity comparison of feature vectors, adaptive dynamic threshold calculation, and abnormal confidence score generation.

8. The intelligent detection system for vehicle barrier collision bars based on deep learning fusion according to claim 7, characterized in that, The linkage control and evidence collection module specifically includes: The camera triggering unit is used to send a trigger command to a preset high-definition network camera after receiving a collision behavior determination signal; The license plate recognition unit is used to process media files captured by the camera and automatically extract license plate information from them; The alarm and recording unit is used to generate structured event records and send alarm messages, while controlling on-site audible and visual alarm devices.

Images