Linkage alarm crime prevention method based on AI dynamic adjustment lighting system

Abstract

The invention relates to the technical field of AI early warning, and discloses a linkage alarm crime prevention method based on an AI dynamic adjustment lighting system, and the method comprises the steps: training and recognizing sounds, including squeal, call-for-help and tense dialogue, made by human beings under dangerous conditions based on an AI algorithm, and when the sounds are recognized, dynamically adjusting the lighting brightness near a sound source, so as to prevent crimes. And then a camera nearby is further utilized to capture a clearer video, an AI algorithm is utilized to further analyze whether illegal behaviors such as stress, trailing and fighting exist or not, if yes, the brightness of surrounding light is further dynamically adjusted, flickering and linkage warning are carried out, potential criminal activities are deterred and early warned, and the purpose of crime prevention is achieved. According to the invention, under the condition of meeting the requirement of energy-saving illumination, the possibility that lawbreakers use related places to implement illegal and criminal behaviors is prevented, and the safety and the user experience are improved.

Description

Technical Field

[0001] The present invention relates to the field of AI early warning technology, and in particular to a linkage alarm crime prevention method based on AI dynamic adjustment of lighting systems. Background Art

[0002] The current energy-saving renovation schemes, especially the renovation schemes for energy-saving lighting in underground garages, mainly include: first, reducing the power consumption and brightness of lighting equipment to ensure basic lighting needs; second, using intelligent lighting tubes, sensing vehicles through radar, etc., dynamically adjusting the lights with the vehicles, and local lighting to achieve the effect of bright lights when the car approaches and dim lights when the car leaves. These schemes can meet basic energy-saving needs, but at the same time, they also make the overall lighting of the basement insufficient. At the same time, due to the small flow of people in the basement, this creates potential conditions for criminals to use such an environment to commit crimes.

[0003] Traditional energy-saving lighting renovation solutions for basements will result in insufficient overall lighting in the basement. At the same time, due to the low traffic volume in the basement, it creates potential conditions for criminals to use such an environment to commit crimes. Summary of the invention

[0004] The present invention provides a linkage alarm crime prevention method based on AI dynamic adjustment of lighting system. While meeting the demand for energy-saving lighting, the present invention prevents the possibility of criminals using related places to commit illegal and criminal acts, thereby improving safety and user experience.

[0005] In a first aspect, the present invention provides a linkage alarm crime prevention method based on AI dynamic adjustment of lighting system, the linkage alarm crime prevention method based on AI dynamic adjustment of lighting system comprising: Collecting an audio data set containing sounds made by humans when facing dangerous situations, and performing discrete Fourier transform and feature extraction on the audio data set to obtain multiple sample audio feature vectors; Inputting the plurality of sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model; Deploy the dangerous sound recognition model to the AI ​​platform, collect the ambient audio data of the parking lot in real time, and analyze and recognize the dangerous sounds on the ambient audio data through the dangerous sound recognition model in the AI ​​platform to obtain the target sound recognition result; When the target sound recognition result is determined to be a dangerous sound, the on-site lighting system of the parking lot is linked to the IoT platform to adjust the light brightness around the sound source, and the corresponding on-site video data is captured by the AI ​​camera; Inputting the on-site video image data into a preset illegal behavior recognition model to analyze and recognize illegal behaviors, and obtaining illegal behavior recognition results; If the illegal behavior identification result is that there is an illegal behavior, the on-site lighting system is linked through the IoT platform to adjust the light brightness and perform flashing warnings and alarms to notify rescue.

[0006] In a second aspect, the present invention provides a linkage alarm crime prevention device based on AI dynamic adjustment lighting system, the linkage alarm crime prevention device based on AI dynamic adjustment lighting system comprises: A collection module, used to collect an audio data set containing sounds made by humans when facing dangerous situations, and perform discrete Fourier transform and feature extraction on the audio data set to obtain multiple sample audio feature vectors; A training module, used for inputting the plurality of sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model; A sound recognition module is used to deploy the dangerous sound recognition model into the AI ​​platform, collect the ambient audio data of the parking lot in real time, analyze and recognize the dangerous sounds of the ambient audio data through the dangerous sound recognition model in the AI ​​platform, and obtain the target sound recognition result; An adjustment module, which is used to adjust the light brightness around the sound source by linking the on-site lighting system of the parking lot through the IoT platform when the target sound recognition result is determined to be a dangerous sound, and to shoot corresponding on-site video picture data through an AI camera; A behavior analysis module, used to input the on-site video data into a preset illegal behavior recognition model to analyze and recognize illegal behaviors, and obtain illegal behavior recognition results; The alarm module is used to, if the illegal behavior identification result is that an illegal behavior exists, link the on-site lighting system through the IoT platform to adjust the light brightness and issue a flashing warning and alarm to notify rescue.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned linkage alarm crime prevention method based on AI dynamic adjustment of the lighting system.

[0008] The fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned linkage alarm crime prevention method based on AI dynamic adjustment of the lighting system.

[0009] In the technical solution provided by the present invention, the present invention is based on AI algorithm training to identify the sounds made by humans in dangerous situations, such as screams, cries for help, tense conversations, etc. When such sounds are identified, the lighting brightness near the sound source can be dynamically adjusted, and then a nearby camera can be used to capture clearer videos, and the AI ​​algorithm can be used to further analyze whether there are illegal acts such as coercion, stalking, and fighting. If so, the surrounding light brightness is further dynamically adjusted and flashed and linked to alarm to deter and warn potential criminal activities, thereby achieving the purpose of preventing crime. While meeting the needs of energy-saving lighting, the present invention also prevents the possibility of criminals using related places to commit illegal and criminal acts, thereby improving safety and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 Schematic diagram of the steps of the linkage alarm crime prevention method based on AI dynamic adjustment of the lighting system in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of the linkage alarm crime prevention device based on AI dynamic adjustment of the lighting system in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] An embodiment of the present invention provides a method for preventing crime by providing a linkage alarm based on an AI-based dynamic adjustment lighting system. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the method for preventing crime by providing a linkage alarm based on AI dynamic adjustment of the lighting system in the embodiment of the present invention includes: Step S1, collecting an audio data set containing sounds made by humans when facing dangerous situations, and performing discrete Fourier transform and feature extraction on the audio data set to obtain multiple sample audio feature vectors; It is understandable that the execution subject of the present invention can be a linkage alarm crime prevention device based on AI dynamic adjustment of lighting system, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0014] Specifically, an audio data set containing sounds made by humans when facing dangerous situations is collected. The audio data set usually includes a variety of typical dangerous sounds such as screams, cries for help, and tense conversations. The audio data set is framed, and the continuous audio data is cut into multiple small time segments to obtain multiple audio frames. Each audio frame is windowed to reduce the impact of spectrum leakage. Commonly used window functions include Hanning window and Hamming window. Fast Fourier transform is performed on the windowed audio frame to convert the time domain signal into a frequency domain signal to obtain multiple audio data represented in the frequency domain. According to the preset Mel filter group, the audio data represented in the frequency domain is filtered. The Mel filter group can simulate the auditory characteristics of the human ear and map the frequency domain data to the Mel frequency scale to obtain multiple Mel spectra. The Mel spectrum can effectively reflect the frequency distribution characteristics in the audio signal, so that the subsequent feature extraction process is more in line with human auditory perception. Logarithmic operations are performed on multiple Mel spectra to compress the dynamic range of the data and enhance the frequency components with smaller amplitudes. Discrete cosine transform is performed on the logarithmic Mel spectrum to obtain Mel frequency cepstral coefficients (MFCC). MFCC is a feature used in audio and speech signal processing. It can effectively capture the frequency characteristics of audio signals and eliminate redundant information of signals to a certain extent. In order to reduce the data dimension and reduce the computational burden, the Mel frequency cepstral coefficients are subjected to dimensionality reduction processing, usually using methods such as principal component analysis. The feature vector after dimensionality reduction is more representative, which improves the training efficiency and prediction accuracy of the model. The feature vector after dimensionality reduction is standardized to eliminate the influence of dimension and obtain the final multiple sample audio feature vectors.

[0015] Step S2: input multiple sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model; Specifically, multiple sample audio feature vectors are subjected to data enhancement processing, and the original data set is expanded by adding noise, changing pitch and speed, etc., to obtain the expanded training data set to enhance the generalization ability of the model. A deep learning model is constructed, where the model is a hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM). The convolutional neural network part extracts spatial features from the audio feature vector, while the long short-term memory network part is used to capture the temporal relationship of audio features, forming a CNN-LSTM hybrid network structure. The convolution part of the CNN-LSTM hybrid network is designed, and a feature extraction subnetwork containing 4 convolution blocks is constructed. Each convolution block consists of 2 two-dimensional convolutional layers, 1 batch normalization layer, and 1 maximum pooling layer. The two-dimensional convolutional layer is used to extract local features, the batch normalization layer is used to accelerate training and stabilize model convergence, and the maximum pooling layer is used to downsample feature maps to reduce the amount of calculation. In order to ensure the nonlinear mapping capability, each convolutional layer uses the ReLU activation function, so that the feature extraction subnetwork can better capture the complex features of audio data. The loop part of the CNN-LSTM hybrid network is designed to construct a time series modeling subnetwork. This part consists of two bidirectional LSTM layers, each containing 256 LSTM units. The bidirectional LSTM can capture the front-end dependency of audio features and enhance the modeling ability of time series information. A global average pooling layer is added to the LSTM layer to reduce the number of model parameters, prevent overfitting and retain important feature information. The feature extraction subnetwork is connected in series with the time series modeling subnetwork, and two fully connected layers are added at the end of the network. The last layer uses a softmax activation function to output the probability of the dangerous sound category corresponding to the audio feature, and the final target CNN-LSTM hybrid network is obtained. In order to ensure the performance of the model, the loss function of the target CNN-LSTM hybrid network is designed, and the cross entropy loss is used as the main loss term, and the L2 regularization term is added to prevent overfitting. Through this design, the training error is effectively reduced while maintaining the generalization ability of the model. The optimizer is configured according to the optimization objective function, and the Adam optimizer is used for training. The Adam optimizer is widely used in the field of deep learning for its adaptive learning rate adjustment ability. The initial learning rate is set to 0.001, and the learning rate is dynamically adjusted through the cosine annealing strategy to improve the efficiency and effect of training. The expanded training data set is divided into training set, validation set and test set according to the preset ratio. During the training process, the target CNN-LSTM hybrid network is iteratively trained according to the optimized training strategy. After each round of training, the model performance is evaluated on the validation set, and the learning rate is dynamically adjusted according to the performance of the validation set to ensure that the model continues to approach the optimal state.After the initial training is completed, the preliminary training model is fine-tuned, and 5 sub-models are trained using the 5-fold cross-validation method. Finally, the prediction results of the sub-models are combined using the soft voting integration method to improve the overall prediction accuracy and stability of the model. Ultimately, a dangerous sound recognition model with superior performance is obtained, which can effectively identify potential dangerous sounds in parking garages.

[0016] Step S3: deploy the dangerous sound recognition model to the AI ​​platform, collect the ambient audio data of the parking lot in real time, analyze and recognize the ambient audio data through the dangerous sound recognition model in the AI ​​platform, and obtain the target sound recognition result; Specifically, the dangerous sound recognition model is deployed in the AI ​​platform to collect and analyze the environmental audio data of the parking garage in real time. The environmental audio data is preprocessed and feature extracted, and the original audio data is converted into a feature vector that can be used for model input to obtain the input feature vector. The input feature vector is input into the first convolution block of the dangerous sound recognition model. The convolution block performs a convolution operation on the input features through the first two-dimensional convolution layer. The convolution layer uses 32 3x3 convolution kernels with a step size of 1 to generate an initial feature map. The ReLU activation function is applied to the initial feature map, and nonlinear mapping is applied to each convolution result to enhance the expressive power of the model. Batch normalization is performed to ensure stability and convergence speed during model training. After processing, the normalized feature map is processed by the second two-dimensional convolution layer. The convolution layer also uses 32 3x3 convolution kernels to extract deeper feature information. The deep feature map is processed with ReLU activation function and batch normalization, and the feature map is downsampled through 2x2 maximum pooling operation with a step size of 2 to obtain the output feature map of the first layer convolution block. The output feature map of the first layer convolution block is input into the second, third and fourth layers of convolution blocks for processing. The structure of each convolution block is the same as that of the first layer convolution block, but the number of convolution kernels increases to 64, 128 and 256 respectively. The second layer convolution block uses 64 convolution kernels to process the output feature map of the first layer convolution block to obtain the output feature map of the second layer convolution block; the third layer convolution block uses 128 convolution kernels to process the output feature map of the second layer convolution block to generate the output feature map of the third layer convolution block; the fourth layer convolution block uses 256 convolution kernels to process the output feature map of the third layer convolution block to generate the target convolution feature map. The target convolution feature map is reshaped to convert it into a time series feature sequence suitable for time series analysis. The time series feature sequence is input into the bidirectional LSTM layer for time series modeling. The bidirectional LSTM layer contains a first bidirectional LSTM layer consisting of 256 LSTM units, which processes both forward and reverse time series information to capture the bidirectional dependency of the audio signal. After processing, the output of the first layer LSTM is dropout to prevent the model from overfitting. The result of the dropout operation is input into the second bidirectional LSTM layer containing 256 LSTM units to generate the LSTM output sequence. The LSTM output sequence is subjected to global average pooling to calculate the average value of the sequence in the time dimension and generate the target feature vector. After the target feature vector is normalized, the normalized feature vector is obtained and input into the fully connected layer network for classification. Through the first fully connected layer, 128 neurons are used to perform a linear transformation on the normalized feature vector to generate the output of the first fully connected layer. The output is subjected to the ReLU activation function and dropout operation to increase the nonlinear expression ability of the model and prevent overfitting.The second fully connected layer performs linear transformation of features, and the number of neurons in this layer is equal to the number of preset dangerous sound categories. The results of the second fully connected layer are processed using the softmax activation function to convert the linear output into the probability distribution of each dangerous sound category. The category with the highest probability is selected as the final target sound recognition result to complete the analysis and recognition of dangerous sounds in the parking lot basement audio data, ensuring that the lighting system can be linked in time for early warning and alarm notification when dangerous situations occur.

[0017] Step S4: When the target sound recognition result is determined to be a dangerous sound, the on-site lighting system of the parking lot is linked to the IoT platform to adjust the light brightness around the sound source, and the corresponding on-site video data is captured by the AI ​​camera; Specifically, when the target sound recognition result is determined to be a dangerous sound, the sound source localization algorithm is triggered to determine the spatial coordinate information of the sound source. Through this algorithm, the specific location of the sound source in the parking garage is located. Based on the acquired spatial coordinate information of the sound source, the three-dimensional spatial model of the parking garage is queried to find the specific location information of the lighting device closest to the sound source. According to the location information of the lighting device, a distance-weighted calculation is performed on each lighting device to obtain the brightness adjustment weight of each device. The closer the distance, the higher the weight of the device, so as to form a stronger lighting effect near the sound source. According to the brightness adjustment weight, the lighting control module in the IoT platform is parameterized, and regional lighting control instructions are generated, and these instructions are sent to the lighting system on site through the communication interface of the IoT platform. The lighting system adjusts the brightness of the lighting devices around the sound source according to the instructions to improve the lighting effect of the sound source area. At the same time, the spatial coordinate information of the sound source is transformed, and the three-dimensional spatial coordinates are mapped to the two-dimensional image coordinate system of the AI ​​camera to determine the specific location of the target area in the camera field of view. According to the image coordinates of the target area, the control parameters of the AI ​​camera's pan / tilt system are calculated to determine the angle the camera needs to rotate and the required zoom parameters so that the camera can accurately aim at the target area. After calculating the pan / tilt rotation angle and zoom parameters, these parameters are sent to the AI ​​camera through the control interface so that the camera automatically adjusts to the required state to ensure that the camera can accurately capture the area where the dangerous sound source is located. After the adjustment is completed, the AI ​​camera is configured for image acquisition, and the acquisition parameters of high resolution and high frame rate are set to ensure that the captured images are clear, smooth, and can provide more detailed information. According to the optimized video acquisition strategy, the AI ​​camera continuously acquires and encodes images of the target area through its image processing module to generate high-quality on-site video data.

[0018] Step S5: inputting the on-site video data into a preset illegal behavior recognition model to analyze and recognize illegal behaviors, and obtaining illegal behavior recognition results; Specifically, the on-site video screen data is preprocessed, including frame extraction, size adjustment and pixel normalization operations, to convert the video frames into standardized input image sequences. The standardized input image sequence is input into the first-layer spatiotemporal convolution block of the 3D convolutional neural network (3D CNN) of the illegal behavior recognition model. The spatiotemporal convolution block uses 64 3x3x3 convolution kernels to extract spatiotemporal features, which can simultaneously capture the temporal relationship and spatial features between video frames. After the convolution operation, the generated initial spatiotemporal feature map is processed by the ReLU activation function to give the model stronger nonlinear expression capabilities. Batch normalization is used to stabilize the model training process, and 3D maximum pooling is performed to reduce the dimension and calculation amount of the feature map to obtain the output feature map of the first-layer spatiotemporal convolution block. The output feature map of the first-layer spatiotemporal convolution block is input into the second, third and fourth-layer spatiotemporal convolution blocks in sequence. The structure of these convolution blocks is similar to the first layer, but the number of convolution kernels increases to 128, 256 and 512 respectively, gradually extracting deeper spatiotemporal features in the video data. Through a layer-by-layer progressive approach, the model can generate multi-scale spatiotemporal feature maps to fully capture the complex dynamic information in the video. The multi-scale spatiotemporal feature maps are globally averaged pooled in the time dimension to compress the temporal information while retaining the key spatial features. The compressed spatial feature maps are input into the residual network module, which consists of 5 residual blocks. Each residual block contains two 2D convolutional layers and a short-circuit connection. Through these residual blocks, the model can extract deep spatial features while retaining the original information through short-circuit connections, enhancing the learning ability and recognition accuracy of the model. The deep spatial features processed by the residual network are globally averaged pooled to compress the feature data and generate a 1024-dimensional feature vector. In order to prevent the model from overfitting, the feature vector is dropout-operated so that some neurons are randomly inactivated during the training process. The 1024-dimensional feature vector is mapped to the number dimension of the illegal behavior category through the fully connected layer to generate the original score of the behavior category. Based on the original score, the model can identify and determine the type of illegal behavior, and finally output the identification results of illegal behaviors including coercion, stalking, and fighting.

[0019] Step S6: If the illegal behavior identification result is that there is an illegal behavior, the on-site lighting system is linked through the IoT platform to adjust the light brightness and issue a flashing warning and alarm notification for rescue.

[0020] Specifically, when the illegal behavior identification result determines that there is an illegal behavior, the linkage response algorithm is triggered and a corresponding linkage response trigger signal is generated. The trigger signal will be used to divide the three-dimensional space model of the parking garage into regions, and divide the entire parking garage into multiple lighting control areas, so as to more flexibly manage and control the lighting equipment in different areas. The lighting equipment in each area is grouped and coded to generate a sub-region lighting equipment control list. The control list assigns independent control parameters to each lighting area, so that the system can perform personalized management of lighting equipment according to the specific conditions of different areas. According to the sub-region lighting equipment control list, the lighting control module in the IoT platform is parameterized and the corresponding sub-region lighting flashing control instructions are generated. After these control instructions are generated, the sub-region lighting flashing control instructions are sent to the on-site lighting system through the communication interface of the IoT platform. According to these instructions, each lighting equipment on the scene will start to perform flashing operations to attract attention and interfere with possible criminal behavior. At the same time, based on the illegal behavior identification results, the preset alarm level model is queried and matched to determine the specific alarm level in the current situation. The alarm level is determined based on the severity of the violation, the location and time of the violation, ensuring the accuracy and effectiveness of the system response. After determining the alarm level, select an alarm information template that matches the current situation from the preset alarm template library. By combining information such as the current time, location, and the type of violation identified, a customized alarm message is generated. The alarm message not only includes a specific description of the violation, but also combines detailed information about the environment and location. Through the message push interface of the IoT platform, the customized alarm message is sent to the rescuers' mobile devices so that they can be notified in time to take necessary countermeasures.

[0021] In the embodiment of the present invention, AI algorithm-based training is used to identify the sounds made by humans in dangerous situations, such as screams, cries for help, tense conversations, etc. When such sounds are identified, the lighting brightness near the sound source can be dynamically adjusted, and then a nearby camera can be used to capture clearer video. The AI ​​algorithm is used to further analyze whether there are illegal acts such as coercion, stalking, and fighting. If so, the surrounding light brightness is further dynamically adjusted and flashed and linked to alarm to deter and warn potential criminal activities, thereby achieving the purpose of preventing crime. While meeting the needs of energy-saving lighting, the present invention prevents the possibility of criminals using related places to commit illegal and criminal acts, thereby improving safety and user experience.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Collect an audio data set containing the sounds made by humans when facing dangerous situations, and divide the audio data set into frames to obtain multiple audio frames. The audio data set includes: screams, cries for help, and tense conversations; Performing windowing processing on each audio frame to obtain multiple windowed audio frames, and performing fast Fourier transform on the multiple windowed audio frames to obtain multiple audio data represented in the frequency domain; Performing filtering processing on the audio data represented in the multiple frequency domains according to a preset Mel filter group to obtain multiple Mel spectra; Performing logarithmic operations on the multiple Mel-frequency spectra to obtain multiple logarithmic Mel-frequency spectra, and performing discrete cosine transform on the multiple logarithmic Mel-frequency spectra to obtain multiple Mel-frequency cepstrum coefficients; Dimensionality reduction processing is performed on multiple Mel-frequency cepstral coefficients respectively to obtain multiple dimensionality-reduced feature vectors, and multiple dimensionality-reduced feature vectors are standardized to obtain multiple sample audio feature vectors.

[0023] Specifically, we collect and process audio data sets containing a variety of typical dangerous sounds such as screams, cries for help, and tense conversations. We divide the collected audio data sets into frames. Framing is to divide the continuous audio signal into smaller time periods, usually with a length of 20 to 40 milliseconds per frame. Audio signals are non-static signals that change over time, and framing can reduce computational complexity while maintaining the temporal characteristics of the signal. Suppose the audio signal is , split it into Each frame signal can be expressed as ,in is the index of the frame. The obtained multiple audio frames It is the basic unit of subsequent signal processing. Windowing is performed on each audio frame to reduce the discontinuity of the signal at the frame boundary, thereby reducing spectrum leakage. Commonly used window functions include Hanning window and Hamming window. Suppose the window function is , then the windowed audio frame can be expressed as By windowing each frame, multiple windowed audio frames are obtained. Perform fast Fourier transform on the windowed audio frame to convert the time domain signal into the frequency domain signal. Fast Fourier transform is an effective algorithm for converting discrete signals from the time domain to the frequency domain. Its formula is: ; in, is the audio data represented in the frequency domain, is the frequency, is the length of the audio frame, is an imaginary unit. Through fast Fourier transform, the frequency domain representation data of each audio frame is obtained, which contains the frequency components of the signal and the corresponding amplitude information. The audio data represented in the frequency domain is filtered according to the preset Mel filter bank. The Mel filter bank is a group of nonlinear filters that can simulate the human ear's perception of different frequencies, thus being closer to the auditory characteristics of the human ear. The center frequency of the Mel filter It can represent: ; The Mel spectrum of each audio frame is obtained by filtering the frequency domain data through a Mel filter bank. The Mel spectrum reflects the energy distribution of the signal on the Mel frequency scale, which helps to identify key sound features in the audio. Logarithmic operations are performed on the Mel spectrum to compress the dynamic range of the data, making the signal components with smaller amplitudes more prominent, thereby improving the effectiveness of the signal in model training. The Mel spectrum after logarithmic processing is called the logarithmic Mel spectrum. The logarithmic Mel spectrum is subjected to discrete cosine transform to generate Mel-frequency cepstrum coefficients. The discrete cosine transform converts the signal from the frequency domain to the cepstrum domain, which can remove redundant information in the spectrum and highlight the characteristics of the signal. The formula for discrete cosine transform is: ; in, It is MFCCs, It is Mel frequency components, is the number of Mel filter banks. The Mel frequency cepstral coefficients are processed for dimensionality reduction to reduce the dimension of the feature vector, reduce the computational burden of the model, and retain the most important feature information. Common dimensionality reduction methods include principal component analysis. The feature vector after dimensionality reduction is more representative and can improve the training efficiency and prediction accuracy of the model. The feature vector after dimensionality reduction is standardized to scale all eigenvalues ​​to the same scale range. Usually, the standardization adjusts the mean of the eigenvalue to 0 and the standard deviation to 1. Through standardization, the scale differences between different features are eliminated, so that the model training is not affected by the feature scale, and multiple sample audio feature vectors are obtained.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Performing data enhancement processing on multiple sample audio feature vectors to obtain an expanded training data set; Construct a deep learning model, which is a hybrid model of convolutional neural network and long short-term memory network, in which the convolution layer is used for feature extraction and the long short-term memory layer is used for time series modeling, and a CNN-LSTM hybrid network structure is obtained; The convolution part of the CNN-LSTM hybrid network is designed, which includes 4 convolution blocks. Each convolution block contains 2 two-dimensional convolution layers, 1 batch normalization layer and 1 maximum pooling layer. The ReLU activation function is used to obtain the feature extraction subnetwork. The loop part of the CNN-LSTM hybrid network is designed to include two bidirectional LSTM layers, each with 256 LSTM units, and a global average pooling layer is added after the LSTM layer to obtain the time series modeling subnetwork; The feature extraction subnetwork and the time series modeling subnetwork are connected in series, and two fully connected layers are added at the end. The last layer uses a softmax activation function to output the probability of dangerous sound categories to obtain the target CNN-LSTM hybrid network. The loss function of the target CNN-LSTM hybrid network is designed, cross entropy loss is used as the main loss term, and L2 regularization term is added to prevent overfitting, and the model optimization objective function is obtained; According to the model optimization objective function, the optimizer of the target CNN-LSTM hybrid network is configured, the Adam optimizer is used, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate to obtain the optimized training strategy; Divide the expanded training data set into training set, validation set and test set according to the preset ratio; According to the optimized training strategy and training set, the target CNN-LSTM hybrid network is iteratively trained. After each round of training, the model performance is evaluated on the validation set, and the learning rate is dynamically adjusted according to the validation set performance to obtain a preliminary training model. The preliminary training model was fine-tuned and integrated, and five sub-models were trained using the 5-fold cross-validation method. The prediction results of these sub-models were combined using the soft voting integration method to finally obtain the dangerous sound recognition model.

[0025] Specifically, data enhancement is performed on multiple sample audio feature vectors to expand the training data set. Data enhancement techniques include adding noise, changing the pitch and speed of the audio, and time scaling, etc. These operations can make the model more robust and generalizable. For example, suppose there is an audio sample of a scream. By time scaling it, multiple audio samples with different lengths are generated to increase the diversity of the training set. The mathematical expression of data enhancement can be expressed as: ; in, is the original audio signal, is the enhanced signal. is the volume adjustment factor, is the time scaling factor, is the added noise. The enhanced data can significantly improve the model's ability to recognize different sounds, and an expanded training data set is obtained. A deep learning model is constructed, which is a hybrid model of a convolutional neural network (CNN) and a long short-term memory network (LSTM). The convolution layer is used to extract audio features, while the LSTM layer is used for time series modeling to capture the temporal dependency of audio signals. When designing the convolution part of the CNN-LSTM hybrid network, a feature extraction subnetwork is constructed. The network contains 4 convolution blocks, each of which contains 2 two-dimensional convolution layers, 1 batch normalization layer, and 1 maximum pooling layer. Each two-dimensional convolution layer uses a 3x3 convolution kernel to extract local spatial features. The batch normalization layer is used to accelerate the training process and stabilize the training of the network, while the maximum pooling layer is used to reduce the dimension of the feature map, thereby reducing the amount of calculation and the risk of overfitting. All convolution layers use the ReLU activation function, and the expression of the ReLU activation function is: ; in, is the feature value output by the convolution layer. The activation function introduces nonlinearity, which enables the model to learn more complex features. Through the design of this structure, a feature extraction subnetwork is formed, which can effectively extract key spatial features from audio data. The loop part of the CNN-LSTM hybrid network is designed to build a time series modeling subnetwork. This subnetwork contains 2 layers of bidirectional LSTM layers, each containing 256 LSTM units. Bidirectional LSTM can capture the forward time dependency of the audio signal, as well as the backward time dependency, to more comprehensively characterize the time series features. The basic unit of LSTM is calculated by the following formula: ; in, For LSTM units at time steps The output, is the state of the output gate, is the cell state, is the hyperbolic tangent activation function. Through bidirectional LSTM, the timing modeling subnetwork can effectively capture the long-term and short-term dependencies in the audio signal. Add a global average pooling layer to the LSTM layer to compress the timing information output by the LSTM layer to obtain a more representative feature vector. The feature extraction subnetwork and the timing modeling subnetwork are connected in series to form a CNN-LSTM hybrid network structure. At the end of the network, add two fully connected layers. The first fully connected layer is used to map the timing features to a lower-dimensional feature space to reduce computational complexity. The output dimension of the second fully connected layer is equal to the number of dangerous sound categories, and the linear output is converted into the probability distribution of each dangerous sound category through the softmax activation function. The expression of the softmax function is: ; in, is the input feature value. The softmax function can map multiple outputs to the probability distribution, thereby determining which category has the highest probability and finally outputting the prediction result of the dangerous sound category. In order to prevent the model from overfitting and ensure the training effect of the model, the loss function of the CNN-LSTM hybrid network is designed, and the cross entropy loss is used as the main loss term, and the L2 regularization term is added. The cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the true distribution. The expression is: ; in, is the true category label, is the predicted class probability, is the regularization coefficient, is the weight parameter of the model. This design effectively prevents overfitting of the model and improves the generalization ability of the model. Configure the optimizer to optimize the training process of the CNN-LSTM hybrid network. The Adam optimizer is used, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate, so that the learning rate can be gradually reduced in the later stage of training, and the global optimal point can be found more accurately. After the model is built and the optimization strategy is determined, the expanded training data set is divided into training set, validation set and test set according to the preset ratio. The training set is used for iterative training of the model, the validation set is used to evaluate the performance of the model after each round of training, and the learning rate is dynamically adjusted to optimize the model, and the test set is used to finally evaluate the generalization ability of the model. In the training stage, the target CNN-LSTM hybrid network is iteratively trained according to the optimization training strategy. After each round of training, the performance of the model on the validation set will be evaluated, and the performance of the model will be gradually improved by adjusting the learning rate and other hyperparameters. After multiple rounds of iterative training, the preliminary training model is finally obtained. The preliminary training model is fine-tuned, and 5 sub-models are trained using the 5-fold cross-validation method. In each cross-validation, the dataset is divided into 5 subsets, each of which is used as a validation set in turn, and the remaining subsets are used as training sets, and finally 5 sub-models are obtained. Through the soft voting ensemble method, the prediction results of these 5 sub-models are combined to form the final prediction result, and a dangerous sound recognition model with higher accuracy and robustness is obtained.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Deploy the dangerous sound recognition model to the AI ​​platform, collect the ambient audio data of the parking garage in real time, pre-process and extract features of the ambient audio data to obtain the input feature vector; The input feature vector is input into the first convolution block of the dangerous sound recognition model. The first convolution block includes: the first two-dimensional convolution layer, using 32 3x3 convolution kernels with a step size of 1, convolution operation is performed on the input feature to obtain the initial feature map; the ReLU activation function is applied to the initial feature map to obtain the activated feature map; the activated feature map is batch normalized to obtain the normalized feature map; the second two-dimensional convolution layer, using 32 3x3 convolution kernels with a step size of 1, convolution operation is performed on the normalized feature map to obtain the deep feature map; the ReLU activation function and batch normalization are applied to the deep feature map again to obtain the processed deep feature map; the processed deep feature map is subjected to a 2x2 maximum pooling operation with a step size of 2 to obtain the output feature map of the first convolution block; The output feature map of the first convolution block is input into the second, third and fourth convolution blocks in sequence. The structure of each convolution block is the same as that of the first convolution block, but the number of convolution kernels is increased to 64, 128 and 256 in sequence to obtain the target convolution feature map. The specific process is as follows: the second convolution block uses 64 convolution kernels to process the output feature map of the first convolution block to obtain the output feature map of the second convolution block; the third convolution block uses 128 convolution kernels to process the output feature map of the second convolution block to obtain the output feature map of the third convolution block; the fourth convolution block uses 256 convolution kernels to process the output feature map of the third convolution block to obtain the target convolution feature map; The target convolution feature map is reshaped to obtain a time series feature sequence; the time series feature sequence is input into the bidirectional LSTM layer, which includes: the first layer of bidirectional LSTM, including 256 LSTM units, processes the forward and reverse time series information at the same time, and obtains the first layer of LSTM output; the first layer of LSTM output is dropout operated; the second layer of bidirectional LSTM, also including 256 LSTM units, processes the output of the first layer of LSTM, and obtains the LSTM output sequence; Perform global average pooling on the LSTM output sequence: calculate the average value of the LSTM output sequence in the time dimension to obtain the target feature vector; standardize the target feature vector to obtain the standardized feature vector; The standardized feature vector is input into the fully connected layer network: the first fully connected layer uses 128 neurons to perform a linear transformation on the standardized feature vector to obtain the output of the first fully connected layer; the output of the first fully connected layer is subjected to the ReLU activation function and dropout operation; the number of neurons in the second fully connected layer is equal to the number of dangerous sound categories, and the features are linearly transformed; the result of the second fully connected layer is subjected to the softmax activation function operation, the linear output is converted into the probability distribution of each dangerous sound category, and the category with the highest probability is selected as the target sound recognition result.

[0027] Specifically, the trained dangerous sound recognition model is integrated on the AI ​​platform. After the model is deployed, the environmental audio data is collected in real time through the microphone installed in the parking garage. In order to ensure that the model can effectively identify potential dangerous sounds, such as screams and cries for help, the original audio data is preprocessed. The continuous audio signal is divided into several audio frames of fixed length by the framing operation of the audio data. Each audio frame is windowed to reduce the influence of the signal boundary effect. The windowed audio frame is subjected to fast Fourier transform to convert the time domain signal into the frequency domain signal. The frequency domain signal is filtered according to the preset Mel filter group to generate a Mel spectrogram. The Mel spectrogram is converted into Mel frequency cepstral coefficients (MFCC) through discrete cosine transform (DCT), and the MFCC is reduced in dimension and normalized to obtain the input feature vector. The standardized input feature vector is input into the first convolution block of the dangerous sound recognition model. The first convolution block contains two two-dimensional convolution layers. The input feature vector is convolved using 32 3x3 convolution kernels. The mathematical expression of the convolution operation is: ; in, is the convolution output, are input features, is the weight of the convolution kernel. Through the convolution operation, the model can extract the initial feature map, reflecting the local spatial features in the input signal. The ReLU activation function is applied to the initial feature map, and the expression of the function is: ; The ReLU activation function suppresses negative values ​​in the feature map to 0, thereby introducing nonlinearity and enabling the model to capture more complex patterns. The activated feature map is batch normalized to stabilize and accelerate the model training process by standardizing the output of each batch. The normalized feature map is convolved for the second time using 32 3x3 convolution kernels to generate a deeper feature map. The ReLU activation function and batch normalization are applied again to ensure the nonlinear characteristics and stability of the deep feature map. The processed deep feature map is subjected to a 2x2 maximum pooling operation to reduce the dimension of the feature map, reduce the computational complexity, and retain the most significant feature information to obtain the output feature map of the first layer convolution block. The output feature map of the first layer convolution block is input into the second, third, and fourth layers of convolution blocks in turn. The structure of each convolution block is similar to the first layer, but the number of convolution kernels increases layer by layer, which are 64, 128, and 256 respectively. The second convolution block uses 64 convolution kernels to process the output feature map of the first layer to generate a new feature map; the third convolution block uses 128 convolution kernels to process the output feature map of the second layer to obtain a higher-level feature map; the fourth convolution block uses 256 convolution kernels to process the output feature map of the third layer to obtain the final target convolution feature map. Through the superposition of multiple layers of convolution blocks, the model can gradually extract high-order spatial features in the audio signal. The target convolution feature map is reshaped to convert the convolution feature map into a time series feature sequence suitable for LSTM layer processing. The reshaped time series feature sequence is input into the bidirectional LSTM layer. The first bidirectional LSTM layer contains 256 LSTM units, which can simultaneously process the forward and reverse time series information of the feature sequence and generate the output of the first LSTM layer. The output of the first LSTM layer is dropout to prevent the model from overfitting by randomly inactivating some neurons. The second bidirectional LSTM layer also contains 256 LSTM units to further process the output of the first LSTM layer to generate a complete LSTM output sequence. Perform a global average pooling operation on the LSTM output sequence, calculate the average value of the output sequence in the time dimension, and obtain the target feature vector. Through global average pooling, the time series information is compressed, making the final feature vector more compact and representative. The target feature vector is standardized to ensure that each feature is on the same scale, thereby preventing certain features from having too much influence on the model prediction. The standardized feature vector is input into the fully connected layer network. The first fully connected layer uses 128 neurons to perform a linear transformation on the standardized feature vector to generate the output of the first fully connected layer. The output is then operated by the ReLU activation function to introduce nonlinear characteristics, and is operated by dropout to reduce the risk of overfitting. The number of neurons in the second fully connected layer is equal to the number of dangerous sound categories, which is used to map the feature vector to a specific dangerous sound category.The model converts the linear output of the second fully connected layer into the probability distribution of each dangerous sound category through the softmax activation function. The expression of the softmax function is:. ; in, is the linear transformation result of the input feature, is the probability of the corresponding category. Through this process, the model can select the category with the highest probability as the target sound recognition result.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: When the target sound recognition result is determined to be a dangerous sound, the sound source localization algorithm is triggered and the spatial coordinate information of the sound source is obtained; According to the spatial coordinate information of the sound source, the three-dimensional spatial model of the parking garage is queried to obtain the location information of the lighting equipment around the sound source, and based on the location information of the lighting equipment, the distance weighted calculation is performed on each lighting equipment to obtain the brightness adjustment weight of the lighting equipment; According to the brightness adjustment weight, the lighting control module in the IoT platform is configured with parameters to obtain the regional lighting control instructions, and the regional lighting control instructions are sent to the on-site lighting system through the communication interface of the IoT platform to adjust the brightness of the lighting equipment around the sound source to enhance the lighting effect; The spatial coordinate information of the sound source is transformed and mapped to the two-dimensional image coordinate system of the AI ​​camera to obtain the image coordinates of the target area. Based on the image coordinates of the target area, the control parameters of the pan / tilt system of the AI ​​camera are calculated to obtain the pan / tilt rotation angle and zoom parameters. The pan / tilt rotation angle and zoom parameters are sent to the AI ​​camera through the control interface, and the AI ​​camera is automatically adjusted to obtain the camera status aimed at the target area. The adjusted AI camera is configured for image acquisition, and the acquisition parameters of high resolution and high frame rate are set to obtain the optimized video acquisition strategy. According to the optimized video acquisition strategy, the target area is continuously acquired and encoded through the image processing module of the AI ​​camera to obtain the on-site video picture data.

[0029] Specifically, when the target sound recognition result is determined to be a dangerous sound, the sound source localization algorithm is triggered to accurately obtain the spatial coordinate information of the sound source. Sound source localization usually relies on multiple microphone arrays, which calculate the location of the sound source by analyzing the time difference (TDOA) of the sound wave arriving at each sensor. Assume that the coordinates of the sound source in three-dimensional space are , the distance between the sound source and each microphone is calculated by the TDOA algorithm, thereby deriving the spatial coordinate information of the sound source. After obtaining the spatial coordinate information of the sound source, the three-dimensional spatial model of the parking garage is queried to determine the location information of the lighting equipment around the sound source. The three-dimensional spatial model of the parking garage is usually pre-established and stored in the system database, including the location, number, and lighting range of all lighting equipment. Based on the coordinate information of the sound source , find the coordinates of all lighting devices near the location ,in Indicates The lighting devices are weighted by distance to determine the lighting contribution of each lighting device to the sound source area. The weighted calculation usually uses a formula inversely proportional to the distance, for example: ; in, It is The weight of each lighting device, is the distance between the lighting device and the sound source, and the calculation formula is: ; Through weighted calculation, the brightness adjustment weight of each lighting device is obtained , the weight is used to adjust the brightness of the lighting equipment so that the equipment close to the sound source has a higher brightness to enhance the lighting effect of the area. According to the brightness adjustment weight, the lighting control module in the IoT platform is parameterized, and regional lighting control instructions are generated to define the brightness level that each lighting device should achieve under different weights, ensuring that the lighting equipment around the sound source can be adjusted according to the preset plan. The generated control instructions are sent to the on-site lighting system through the communication interface of the IoT platform. After receiving the instructions, the on-site lighting system adjusts the brightness of each related lighting device in real time to optimize the lighting environment near the sound source. At the same time, the spatial coordinate information of the sound source is transformed and mapped to the two-dimensional image coordinate system of the AI ​​camera. Assume that the three-dimensional spatial coordinate system of the parking garage is , the two-dimensional image coordinate system of the Al camera is , the coordinate transformation can be achieved through the camera's internal and external parameter matrix. The transformation formula is usually: ; in, is the intrinsic parameter matrix of the camera, is the rotation matrix, is the translation vector. By transforming, the three-dimensional space coordinates Convert to 2D image coordinates in the camera's field of view , determine the specific location of the sound source in the camera image. According to the image coordinates of the target area , calculate the control parameters of the AI ​​camera's pan / tilt system and determine the camera's rotation angle and zoom parameters. and vertical rotation angle , the camera can be accurately aimed at the target area to ensure that the sound source area is within the monitoring range of the camera. The calculation of the rotation angle can be achieved through trigonometric function relationships, for example: ; in, and are the coordinates of the center of the image, is the focal length of the camera. The zoom parameter is adjusted according to the distance of the sound source and the focal length of the camera to optimize the clarity of the monitoring image. The pan / tilt rotation angle and zoom parameters are sent to the AI ​​camera through the control interface, and the camera automatically adjusts to the state of aiming at the target area. After the pan / tilt adjustment is completed, the image acquisition configuration of the camera is optimized, including setting the acquisition parameters of high resolution and high frame rate to ensure that the camera can capture clear and smooth live video images. The optimized configuration aims to maximize the effect of video surveillance so that the system can accurately record and analyze the details of dangerous events. According to the optimized video acquisition strategy, the target area is continuously acquired and encoded through the image processing module of the AI ​​camera to generate high-quality live video data.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Perform frame extraction, size adjustment and pixel normalization on the live video data to obtain a standardized input image sequence; The standardized input image sequence is input into the first-layer spatiotemporal convolution block of the 3D convolutional neural network in the illegal behavior recognition model, and 64 3x3x3 convolution kernels are used to extract spatiotemporal features to obtain the initial spatiotemporal feature map. The initial spatiotemporal feature map is subjected to ReLU activation function and batch normalization processing, and a 3D maximum pooling operation is performed to obtain the output feature map of the first-layer spatiotemporal convolution block; The output feature map of the first-layer spatiotemporal convolution block is sequentially input into the second-layer spatiotemporal convolution block, the third-layer spatiotemporal convolution block, and the fourth-layer spatiotemporal convolution block. The number of convolution kernels in each layer of spatiotemporal convolution block increases, which are 128, 256, and 512 respectively, to obtain a multi-scale spatiotemporal feature map; Perform global average pooling on the multi-scale spatiotemporal feature map in the time dimension to obtain a compressed spatial feature map, and input the compressed spatial feature map into the residual network module. The residual network module contains 5 residual blocks, each of which contains two 2D convolutional layers and a short-circuit connection to obtain deep spatial features; The deep spatial features are globally averaged pooled and transformed into a fully connected layer to obtain a 1024-dimensional feature vector, and the 1024-dimensional feature vector is dropout-operated. The 1024-dimensional feature vector is then mapped to the dimension of the number of illegal behavior categories through a fully connected layer to obtain the original scores of the behavior categories. The illegal behavior recognition results are generated based on the original scores of the behavior categories, and the illegal behavior recognition results include coercion, stalking, and fighting.

[0031] Specifically, the raw video data is preprocessed, including steps such as frame extraction, resizing, and pixel normalization. Frame extraction is to extract continuous frames in the video stream into a series of independent static images at fixed time intervals. Assuming that the video frame rate is 30 frames per second, through frame extraction, a frame is extracted every few frames, thereby reducing the amount of calculation and ensuring that each image contains sufficient information. The size of each extracted frame image is resized, and the image is usually scaled to a fixed size, such as 224x224 pixels. The size of the input data is unified to adapt to the subsequent deep learning model. The pixel values ​​of each image are normalized and the pixel values ​​are scaled to the [0, 1] interval to improve the training stability and convergence speed of the model, and a standardized input image sequence is obtained. The standardized image sequence is input into the first layer of the spatiotemporal convolution block of the 3D convolutional neural network (3DCNN) in the illegal behavior recognition model. The 3D convolution layer is different from the traditional 2D convolution layer. It not only performs convolution in the spatial dimension, but also in the temporal dimension, and can simultaneously capture the spatiotemporal features in the video sequence. Assume that the input image sequence is , whose size is ,in is the number of time frames, and are the height and width of the image, is the number of channels. Use 64 3x3x3 convolution kernels to perform convolution operations on the input image sequence to generate the initial spatiotemporal feature map The formula for the convolution operation is ; in, is the weight of the convolution kernel, is in space-time The convolution process can extract the features of local spatiotemporal regions in the video sequence. The initial feature map is processed by ReLU activation function, nonlinearity is introduced, negative values ​​are filtered out, and they are set to 0. The activated feature map is processed by batch normalization to stabilize the training process and avoid gradient disappearance or explosion. The processed feature map is subjected to 3D maximum pooling operation. The role of pooling is to downsample the feature map to reduce computational complexity and reduce the risk of overfitting. The operation formula of maximum pooling is: ; in, is the feature map after pooling. After processing, the output feature map of the first layer of spatiotemporal convolutional block is obtained. The output feature map of the first layer of spatiotemporal convolutional block is input into the second, third and fourth layers of spatiotemporal convolutional blocks in sequence. The structure of these convolutional blocks is similar to that of the first layer, but the number of convolution kernels in each layer increases, namely 128, 256 and 512 respectively. Through the layer-by-layer processing of these convolutional blocks, the model can extract richer multi-scale spatiotemporal features and form a multi-scale spatiotemporal feature map. The multi-scale spatiotemporal feature map is subjected to global average pooling operation in the time dimension. The average value of the feature map is calculated in the time dimension to obtain a compressed spatial feature map. The compressed spatial feature map is input into the residual network module. The residual network module contains 5 residual blocks, each of which consists of two 2D convolutional layers and a short-circuit connection. The residual block introduces a skip connection through a short-circuit connection to prevent the information from gradually disappearing in the deep network and ensure that the features can be transmitted more effectively. Through the residual network module, deep spatial features are extracted. The deep spatial features are globally averaged and pooled, and then transformed through a fully connected layer to generate a 1024-dimensional feature vector. To prevent overfitting, a dropout operation is performed on the feature vector after the fully connected layer to randomly discard a portion of neurons to improve the generalization ability of the model. The 1024-dimensional feature vector is input into a fully connected layer and mapped to the dimension of the violation category. The output of the fully connected layer is converted into a probability distribution through the softmax activation function. The formula for softmax is: ; in, It is The raw scores of the categories, is its corresponding probability. Through this process, the model can generate a raw score for each violation category, and calculate the probability distribution of each category based on these scores, and select the category with the highest probability as the final violation identification result. The violation identification results may include categories such as coercion, stalking, and fighting.

[0032] In a specific embodiment, the process of executing step S6 may specifically include the following steps: If the illegal behavior identification result is that there is an illegal behavior, the linkage response algorithm is triggered and a linkage response trigger signal is generated. According to the linkage response trigger signal, the three-dimensional space model of the parking garage is divided into regions to obtain multiple lighting control areas; Based on multiple lighting control areas, the lighting devices in each area are grouped and coded to obtain a regional lighting device control list. According to the regional lighting device control list, the lighting control module in the IoT platform is parameterized to obtain regional lighting flicker control instructions. The regional lighting flickering control instructions are sent to the on-site lighting system through the communication interface of the IoT platform to control the flickering of lighting equipment in each area. Based on the illegal behavior identification results, the preset alarm level model is queried and matched to obtain the alarm level of the current situation. According to the alarm level, select the corresponding alarm information template from the preset alarm template library, and combine it with the current time and location information to obtain customized alarm information; The customized alarm information is sent to the rescue personnel's mobile devices through the message push interface of the IoT platform to warn and notify the rescue.

[0033] Specifically, if the result of the illegal behavior identification is that there is an illegal behavior, the linkage response algorithm is triggered and a corresponding linkage response trigger signal is generated. According to the trigger signal, the three-dimensional space model of the parking garage is analyzed and divided into regions. The three-dimensional space model of the parking garage is usually established in advance, and the geographical information and facility locations of all regions are recorded. Based on the model, the specific location where the illegal behavior occurred is identified, and the parking lot is divided into multiple lighting control areas. These areas can be divided according to different spatial locations, lighting requirements and other relevant factors, so as to ensure that the lighting equipment in each area is managed in a refined manner in an emergency. Based on the lighting control area, the lighting equipment in each area is grouped and coded. The purpose of grouping is to uniformly manage the lighting equipment in the same area so that the control instructions can be accurate to a specific area, while the coding assigns a unique identifier to each device to ensure the accuracy of the communication and control process. Through grouping and coding, a regional lighting equipment control list is generated to record the lighting equipment in each area and its corresponding control parameters. According to the regional lighting equipment control list, the lighting control module in the IoT platform is parameterized. The configured parameters include information such as the lighting brightness, flashing frequency, and flashing mode of each area to ensure that the lighting equipment can respond according to the predetermined plan after receiving the control command. After the configuration is completed, the regional lighting flashing control command is generated, including the operation details of all lighting equipment in each area. The regional lighting flashing control command is sent to the on-site lighting system through the communication interface of the IoT platform. After receiving the command, the on-site lighting system controls the flashing of the lighting equipment in each relevant area. At the same time, based on the results of illegal behavior identification, the preset alarm level model is queried and matched to determine the specific alarm level in the current situation. The alarm level is usually determined based on multi-dimensional factors such as the severity, time, and location of the illegal behavior, and may be divided into multiple levels such as low, medium, high, and emergency. The process of matching the alarm level can be carried out through a formula, for example: ; in, Represents the alarm level. is the severity of the offense, is the time factor of the event. It is the sensitivity of location. , , γ are the weight coefficients of each factor. The alarm level of the current situation is calculated through the formula, so as to determine the level of response measures to be taken. After determining the alarm level, select an alarm information template that matches the current situation from the preset alarm template library. The alarm template library usually contains standardized alarm information in a variety of situations, such as early warning information for different types of illegal behaviors (such as coercion, stalking, fighting) or different levels of urgency. The template provides the basic framework and format of the alarm information. Combined with the current time, location and other specific information, the selected alarm template is customized. For example, the time of the incident, the specific location (such as a certain area in the parking lot), and the type of illegal behavior identified are clearly marked in the alarm information. The customized alarm information is sent to the mobile device of the rescue personnel through the message push interface of the IoT platform. The message push process can be implemented through the platform's API interface to deliver the alarm information to the relevant personnel in real time to ensure that they can receive the alarm and respond in the first time. For example, when the system identifies a fight in a certain area of ​​the parking lot, an alarm message will be immediately sent to the security guard or manager's mobile phone. The message will include the specific location of the fight, the time of occurrence, and the response measures recommended by the system.

[0034] The above describes the method for preventing crime by using a linkage alarm based on the dynamic adjustment of the lighting system by AI in the embodiment of the present invention. The following describes the device for preventing crime by using a linkage alarm based on the dynamic adjustment of the lighting system by AI in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a linkage alarm crime prevention device based on AI dynamic adjustment of lighting system includes: The acquisition module is used to acquire an audio data set containing sounds made by humans when facing dangerous situations, and to perform discrete Fourier transform and feature extraction on the audio data set to obtain multiple sample audio feature vectors; A training module, used to input multiple sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model; The sound recognition module is used to deploy the dangerous sound recognition model to the AI ​​platform and collect the environmental audio data of the parking garage in real time. The dangerous sound recognition model in the AI ​​platform is used to analyze and recognize the environmental audio data to obtain the target sound recognition result. The adjustment module is used to adjust the brightness of the lights around the sound source through the IoT platform and the parking lot's on-site lighting system when the target sound recognition result is determined to be a dangerous sound, and to capture the corresponding on-site video data through the AI ​​camera; The behavior analysis module is used to input the on-site video data into the preset illegal behavior recognition model to analyze and recognize illegal behaviors and obtain illegal behavior recognition results; The alarm module is used to link the on-site lighting system through the IoT platform to adjust the light brightness and issue flashing warnings and alarms to notify rescue if the illegal behavior identification result is that an illegal behavior has occurred.

[0035] Through the collaborative cooperation of the above-mentioned components, the present invention is based on AI algorithm training to identify the sounds made by humans in dangerous situations, such as screams, cries for help, tense conversations, etc. When such sounds are identified, the lighting brightness near the sound source can be dynamically adjusted, and then a nearby camera can be used to capture clearer video, and the AI ​​algorithm can be used to further analyze whether there are illegal acts such as coercion, stalking, and fighting. If so, the surrounding light brightness is further dynamically adjusted and flashed and linked to alarm to deter and warn potential criminal activities, thereby achieving the purpose of preventing crime. While meeting the needs of energy-saving lighting, the present invention also prevents the possibility of criminals using related places to commit illegal and criminal acts, thereby improving safety and user experience.

[0036] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the linkage alarm crime prevention method based on AI dynamic adjustment of the lighting system in the above-mentioned embodiments.

[0037] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer executes the steps of the linkage alarm crime prevention method based on AI dynamic adjustment of the lighting system.

[0038] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0039] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0040] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for preventing crime by using a linkage alarm based on AI to dynamically adjust the lighting system, characterized in that: The method comprises: Collecting an audio data set containing sounds made by humans when facing dangerous situations, and performing discrete Fourier transform and feature extraction on the audio data set to obtain multiple sample audio feature vectors; Inputting the plurality of sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model; Deploy the dangerous sound recognition model to the AI ​​platform, collect the ambient audio data of the parking lot in real time, and analyze and recognize the dangerous sounds on the ambient audio data through the dangerous sound recognition model in the AI ​​platform to obtain the target sound recognition result; When the target sound recognition result is determined to be a dangerous sound, the on-site lighting system of the parking lot is linked to the IoT platform to adjust the light brightness around the sound source, and the corresponding on-site video data is captured by the AI ​​camera; Inputting the on-site video image data into a preset illegal behavior recognition model to analyze and recognize illegal behaviors, and obtaining illegal behavior recognition results; If the illegal behavior identification result is that there is an illegal behavior, the on-site lighting system is linked through the IoT platform to adjust the light brightness and perform flashing warnings and alarms to notify rescue.

2. The method for preventing crime by using a linkage alarm based on AI dynamic lighting system adjustment according to claim 1 is characterized in that: The acquisition includes an audio data set of sounds made by humans when facing dangerous situations, and the audio data set is subjected to discrete Fourier transform and feature extraction to obtain a plurality of sample audio feature vectors, including: Collecting an audio data set containing sounds made by humans when facing dangerous situations, and dividing the audio data set into frames to obtain multiple audio frames, wherein the audio data set includes: screams, cries for help, and tense conversations; Performing windowing processing on each of the audio frames to obtain a plurality of windowed audio frames, and performing fast Fourier transform on the plurality of windowed audio frames to obtain a plurality of audio data represented in the frequency domain; Performing filtering processing on the plurality of audio data represented in the frequency domains according to a preset Mel filter group to obtain a plurality of Mel spectra; Performing logarithmic operations on the multiple Mel-frequency spectra to obtain multiple logarithmic Mel-frequency spectra, and performing discrete cosine transform on the multiple logarithmic Mel-frequency spectra to obtain multiple Mel-frequency cepstrum coefficients; Dimensionality reduction processing is performed on the multiple Mel-frequency cepstral coefficients respectively to obtain multiple dimension-reduced feature vectors, and standardization processing is performed on the multiple dimension-reduced feature vectors respectively to obtain multiple sample audio feature vectors.

3. The method for preventing crime by using a linkage alarm based on AI dynamic lighting system adjustment according to claim 1, characterized in that: The step of inputting the plurality of sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model comprises: Performing data enhancement processing on the multiple sample audio feature vectors to obtain an expanded training data set; Constructing a deep learning model, wherein the deep learning model is a hybrid model of a convolutional neural network and a long short-term memory network, wherein the convolutional layer is used for feature extraction and the long short-term memory layer is used for time series modeling, thereby obtaining a CNN-LSTM hybrid network structure; The convolution part of the CNN-LSTM hybrid network is designed to include 4 convolution blocks, each of which includes 2 two-dimensional convolution layers, 1 batch normalization layer and 1 maximum pooling layer, and a ReLU activation function is used to obtain a feature extraction subnetwork; The loop part of the CNN-LSTM hybrid network is designed to include two bidirectional LSTM layers, each with 256 LSTM units, and a global average pooling layer is added after the LSTM layer to obtain a time series modeling subnetwork; The feature extraction subnetwork and the time series modeling subnetwork are connected in series, and two fully connected layers are added at the end. The last layer uses a softmax activation function to output the probability of dangerous sound categories, thereby obtaining a target CNN-LSTM hybrid network. A loss function is designed for the target CNN-LSTM hybrid network, cross entropy loss is used as the main loss term, and an L2 regularization term is added to prevent overfitting, thereby obtaining a model optimization objective function; According to the model optimization objective function, the target CNN-LSTM hybrid network is optimized, the Adam optimizer is used, the initial learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate to obtain an optimized training strategy; Dividing the expanded training data set into a training set, a validation set, and a test set according to a preset ratio; According to the optimization training strategy and the training set, the target CNN-LSTM hybrid network is iteratively trained, the model performance is evaluated on the validation set after each round of training, and the learning rate is dynamically adjusted according to the validation set performance to obtain a preliminary training model; The preliminary training model is fine-tuned and integrated, 5 sub-models are trained using a 5-fold cross-validation method, and the prediction results of these sub-models are combined using a soft voting integration method to finally obtain a dangerous sound recognition model.

4. The method for preventing crime by using a linkage alarm based on AI dynamic lighting system adjustment according to claim 1, characterized in that: The dangerous sound recognition model is deployed in the AI ​​platform, and the environmental audio data of the parking lot is collected in real time. The dangerous sound recognition model in the AI ​​platform is used to analyze and recognize the environmental audio data to obtain the target sound recognition result, including: Deploy the dangerous sound recognition model to the AI ​​platform, collect the ambient audio data of the parking garage in real time, pre-process and extract features of the ambient audio data, and obtain an input feature vector; Input the input feature vector into the first convolution block of the dangerous sound recognition model, the first convolution block comprising: a first two-dimensional convolution layer, using 32 3x3 convolution kernels with a step size of 1, performing a convolution operation on the input feature to obtain an initial feature map; applying a ReLU activation function to the initial feature map to obtain an activated feature map; performing batch normalization on the activated feature map to obtain a normalized feature map; a second two-dimensional convolution layer, using 32 3x3 convolution kernels with a step size of 1, performing a convolution operation on the normalized feature map to obtain a deep feature map; applying a ReLU activation function and batch normalization to the deep feature map again to obtain a processed deep feature map; performing a 2x2 maximum pooling operation on the processed deep feature map with a step size of 2 to obtain an output feature map of the first convolution block; The output feature map of the first-layer convolution block is sequentially input into the second-layer convolution block, the third-layer convolution block and the fourth-layer convolution block. The structure of each convolution block is the same as that of the first-layer convolution block, but the number of convolution kernels is increased to 64, 128 and 256 respectively to obtain the target convolution feature map. The specific process is as follows: the second-layer convolution block uses 64 convolution kernels to process the output feature map of the first-layer convolution block to obtain the output feature map of the second-layer convolution block; the third-layer convolution block uses 128 convolution kernels to process the output feature map of the second-layer convolution block to obtain the output feature map of the third-layer convolution block; the fourth-layer convolution block uses 256 convolution kernels to process the output feature map of the third-layer convolution block to obtain the target convolution feature map; The target convolution feature map is reshaped to obtain a time series feature sequence; the time series feature sequence is input into a bidirectional LSTM layer, wherein the bidirectional LSTM layer comprises: a first layer of bidirectional LSTM, comprising 256 LSTM units, processing forward and reverse time series information at the same time, and obtaining a first layer of LSTM output; a dropout operation is performed on the first layer of LSTM output; a second layer of bidirectional LSTM, also comprising 256 LSTM units, processing the output of the first layer of LSTM, and obtaining an LSTM output sequence; Performing a global average pooling operation on the LSTM output sequence: calculating the average value of the LSTM output sequence in the time dimension to obtain a target feature vector; performing a normalization process on the target feature vector to obtain a normalized feature vector; The standardized feature vector is input into the fully connected layer network: the first fully connected layer uses 128 neurons to perform a linear transformation on the standardized feature vector to obtain the output of the first fully connected layer; the output of the first fully connected layer is subjected to a ReLU activation function operation and a dropout operation; the number of neurons in the second fully connected layer is equal to the number of dangerous sound categories, and a feature linear transformation is performed; the result of the second fully connected layer is subjected to a softmax activation function operation, the linear output is converted into a probability distribution of each dangerous sound category, and the category with the highest probability is selected as the target sound recognition result.

5. The method for preventing crime by using a linkage alarm based on AI dynamic lighting system adjustment according to claim 1, characterized in that: When the target sound recognition result is determined to be a dangerous sound, the on-site lighting system of the parking lot is linked to the IoT platform to adjust the light brightness around the sound source, and the corresponding on-site video image data is captured by the AI ​​camera, including: When the target sound recognition result is determined to be a dangerous sound, a sound source localization algorithm is triggered and the spatial coordinate information of the sound source is obtained; According to the spatial coordinate information of the sound source, the three-dimensional spatial model of the parking garage is queried to obtain the position information of the lighting equipment around the sound source, and based on the lighting equipment position information, a distance weighted calculation is performed on each lighting equipment to obtain the brightness adjustment weight of the lighting equipment; According to the brightness adjustment weight, the lighting control module in the IoT platform is configured with parameters to obtain a regional lighting control instruction, and the regional lighting control instruction is sent to the on-site lighting system through the communication interface of the IoT platform to adjust the brightness of the lighting equipment around the sound source to enhance the lighting effect; The spatial coordinate information of the sound source is transformed and mapped into the two-dimensional image coordinate system of the AI ​​camera to obtain the image coordinates of the target area, and the control parameters of the pan / tilt system of the AI ​​camera are calculated according to the image coordinates of the target area to obtain the pan / tilt rotation angle and zoom parameters; The pan / tilt rotation angle and zoom parameters are sent to the AI ​​camera through the control interface, the AI ​​camera is automatically adjusted to obtain the camera state aimed at the target area, and the adjusted AI camera is configured for image acquisition, and acquisition parameters of high resolution and high frame rate are set to obtain an optimized video acquisition strategy; According to the optimized video acquisition strategy, the target area is continuously acquired and encoded through the image processing module of the AI ​​camera to obtain on-site video picture data.

6. The method for preventing crime by using a linkage alarm based on AI dynamic lighting system adjustment according to claim 1, characterized in that: The step of inputting the on-site video data into a preset illegal behavior recognition model to analyze and recognize illegal behaviors and obtain illegal behavior recognition results includes: Performing frame extraction, size adjustment and pixel normalization on the live video picture data to obtain a standardized input image sequence; Input the standardized input image sequence into the first-layer spatiotemporal convolution block of the 3D convolutional neural network in the illegal behavior recognition model, use 64 3x3x3 convolution kernels to extract spatiotemporal features to obtain an initial spatiotemporal feature map, perform ReLU activation function and batch normalization processing on the initial spatiotemporal feature map, and perform 3D maximum pooling operation to obtain an output feature map of the first-layer spatiotemporal convolution block; The output feature map of the first-layer spatiotemporal convolution block is sequentially input into the second-layer spatiotemporal convolution block, the third-layer spatiotemporal convolution block and the fourth-layer spatiotemporal convolution block, and the number of convolution kernels of each layer of spatiotemporal convolution block increases, which are 128, 256 and 512 respectively, to obtain a multi-scale spatiotemporal feature map; Performing global average pooling on the multi-scale spatiotemporal feature map in the time dimension to obtain a compressed spatial feature map, and inputting the compressed spatial feature map into a residual network module, wherein the residual network module includes 5 residual blocks, each residual block includes two 2D convolutional layers and a short-circuit connection, to obtain deep spatial features; The deep spatial features are subjected to global average pooling and fully connected layer transformation to obtain a 1024-dimensional feature vector, and a dropout operation is performed on the 1024-dimensional feature vector. The 1024-dimensional feature vector is then mapped to the dimension of the number of illegal behavior categories through a fully connected layer to obtain the original scores of the behavior categories, and illegal behavior recognition results are generated according to the original scores of the behavior categories, and the illegal behavior recognition results include coercion, stalking, and fighting.

7. The method for preventing crime by using a linkage alarm based on AI dynamic lighting system adjustment according to claim 1, characterized in that: If the illegal behavior identification result is that there is an illegal behavior, the on-site lighting system is linked to the IoT platform to adjust the light brightness and perform flashing warning and alarm notification for rescue, including: If the illegal behavior identification result is that there is an illegal behavior, the linkage response algorithm is triggered and a linkage response trigger signal is generated. According to the linkage response trigger signal, the three-dimensional space model of the parking garage is divided into regions to obtain multiple lighting control areas; Based on the multiple lighting control areas, lighting devices in each area are grouped and encoded to obtain a sub-area lighting device control list, and according to the sub-area lighting device control list, the lighting control module in the IoT platform is parameterized to obtain a sub-area lighting flicker control instruction; The regional lighting flickering control command is sent to the on-site lighting system through the communication interface of the IoT platform to control the flickering of lighting equipment in each area, and based on the illegal behavior identification result, the preset alarm level model is queried and matched to obtain the alarm level of the current situation; According to the alarm level, a corresponding alarm information template is selected from a preset alarm template library, and customized alarm information is obtained by combining the current time and location information; The customized alarm information is sent to the rescue personnel's mobile device through the message push interface of the IoT platform for alarm notification and rescue.

8. A linkage alarm crime prevention device based on AI dynamic adjustment lighting system, characterized in that: The device is used to execute the linkage alarm crime prevention method based on AI dynamic adjustment lighting system according to any one of claims 1 to 7, and comprises: A collection module, used to collect an audio data set containing sounds made by humans when facing dangerous situations, and perform discrete Fourier transform and feature extraction on the audio data set to obtain multiple sample audio feature vectors; A training module, used for inputting the plurality of sample audio feature vectors into a preset deep learning model for model training to obtain a dangerous sound recognition model; A sound recognition module is used to deploy the dangerous sound recognition model into the AI ​​platform, collect the ambient audio data of the parking lot in real time, analyze and recognize the dangerous sound of the ambient audio data through the dangerous sound recognition model in the AI ​​platform, and obtain the target sound recognition result; An adjustment module, which is used to adjust the light brightness around the sound source by linking the on-site lighting system of the parking garage through the IoT platform when the target sound recognition result is determined to be a dangerous sound, and to shoot corresponding on-site video picture data through the AI ​​camera; A behavior analysis module, used to input the on-site video data into a preset illegal behavior recognition model to analyze and recognize illegal behaviors, and obtain illegal behavior recognition results; The alarm module is used to, if the illegal behavior identification result is that an illegal behavior exists, link the on-site lighting system through the IoT platform to adjust the light brightness and issue a flashing warning and alarm to notify rescue.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the linkage alarm crime prevention method based on AI dynamic adjustment of lighting system as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the linkage alarm crime prevention method based on AI dynamic adjustment of lighting system as described in any one of claims 1 to 7.

Images