Vehicle automatic sound production control method based on infrared recognition

Abstract

The invention provides a vehicle automatic sound production control method based on infrared recognition, which comprises the following steps: acquiring an infrared image in a preset range in front of a vehicle, recognizing a heat source target in the image, and judging whether the heat source target is an animal or a human according to the shape, size and temperature distribution characteristics of the heat source target; the prediction unit is used for predicting the moving track of the animal and the human in a period of time in the future according to typical moving track modes of the animal and the human so as to obtain predicted moving position and speed information, and transmitting the predicted information to the acoustic control unit; the acoustic control unit dynamically adjusts the sound pressure level and the frequency of the loudspeaker according to the received predicted moving information and the target type so as to generate an acoustic warning mode matched with the predicted moving position, the predicted moving speed and the target type; when the acoustic signal is sent out, the actual moving track of an animal or a person is continuously tracked, and the track prediction method is optimized by comparing the deviation between the actual track and the predicted track, so that the optimal avoidance effect is achieved.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to an automatic vehicle sound control method based on infrared recognition. Background Art

[0002] In an urban environment, when animals or people may collide with a vehicle, the acoustic control unit needs to adjust the horn sound pressure level in real time according to the movement trajectory of the animals and people to achieve the best avoidance effect. However, urban residents usually do not like sudden loud horn sounds and are easily frightened. When the system identifies the heat source characteristics of animals and people, the acoustic control unit needs to adjust the horn sound pressure level in real time according to the movement trajectory of animals and people to achieve the best avoidance effect. This requires the acoustic control unit to accurately predict the movement trajectory of animals and people, and adjust the horn sound pressure level in advance according to the prediction results, so that the change in sound pressure level matches the movement trajectory of animals and people. However, the movement trajectory of animals and people often has great uncertainty and randomness, and different types of animals and people have different sensitivities to sound, which brings great challenges to the prediction and adjustment of the acoustic control unit. At the same time, too high a sound pressure level may have a frightening effect on animals and people, causing them to behave unpredictably and even cause traffic accidents; while too low a sound pressure level may not attract the attention of animals and people, and fail to achieve the expected avoidance effect. Therefore, how to dynamically adjust the sound pressure level according to the movement trajectories of animals and humans and ultimately achieve the best avoidance effect is a key technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention provides a vehicle automatic sound control method based on infrared recognition, which mainly includes:

[0004] Obtain infrared images within a predetermined range in front of the vehicle, identify heat source targets in the image, and determine whether the heat source targets are animals or humans based on their shape, size, and temperature distribution characteristics;

[0005] For the identified heat source targets of animals or people, their movement trajectories are tracked in real time, and the movement trajectory data within a predetermined time is obtained. The trajectory data is analyzed through the trajectory clustering algorithm to obtain the movement trajectory pattern of the animals or people;

[0006] According to the movement trajectory pattern of the animal or person, predict the movement trajectory within the predetermined time to obtain predicted movement position and speed information, and transmit the predicted information to the acoustic control unit;

[0007] The acoustic control unit dynamically adjusts the sound pressure level and frequency of the horn according to the received predicted movement information and target category to generate an acoustic warning method that matches the predicted movement position, speed and target type;

[0008] If the target is identified as a pedestrian, the acoustic control unit inputs the target type, predicted trajectory, and speed parameters into the pre-trained decision neural network, learns based on historical data, generates a warning strategy for pedestrians, and obtains the real-time frequency and volume adjustment curve of the prompt sound. The acoustic control unit generates a dynamic prompt sound signal based on this, which changes continuously as the target approaches;

[0009] If the target is identified as an animal or a vehicle, the acoustic control unit inputs the target parameters into the pre-trained reinforcement learning model to obtain the optimal warning time point, volume step adjustment amplitude and frequency. The acoustic control unit generates a horn signal in real time and adjusts it adaptively according to the target's reaction until the safe avoidance goal is achieved.

[0010] While sending out acoustic signals, the actual movement trajectory of animals or people is continuously tracked. By comparing the deviation between the actual trajectory and the predicted trajectory, the trajectory prediction method is optimized to achieve the best avoidance effect.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a method for automatic vehicle sound control based on infrared recognition. The method uses an infrared thermal imager to obtain an image in front of the vehicle, uses image segmentation and target recognition algorithms to determine whether the heat source target is an animal or a person, and tracks its movement trajectory in real time. Through trajectory clustering and prediction algorithms, the future movement position and speed of the target can be predicted. According to the prediction information, the acoustic control unit dynamically adjusts the sound pressure level and frequency of the horn to generate a targeted acoustic warning strategy. For pedestrians, a decision neural network is used to generate the optimal warning strategy; for animals or vehicles, a reinforcement learning model is used to determine the best warning plan. It can also adaptively adjust according to the actual response of the target, and optimize the trajectory prediction model in real time through a machine learning algorithm. This intelligent acoustic warning method can effectively improve vehicle driving safety and reduce the incidence of traffic accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a vehicle automatic sound control method based on infrared recognition. DETAILED DESCRIPTION

[0014] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0015] like Figure 1 In this embodiment, a vehicle automatic sound control method based on infrared recognition may specifically include:

[0016] Step S101, acquiring an infrared image within a predetermined range in front of the vehicle, identifying a heat source target in the image, and determining whether the heat source target is an animal or a person based on the shape, size, and temperature distribution characteristics of the heat source target.

[0017] An infrared thermal imager is used to obtain an infrared image in front of a vehicle, and a temperature correction curve is established by calibrating the temperature of a blackbody radiation source and the actual temperature. The infrared image is corrected for brightness and temperature to obtain a first infrared temperature map; the image quality level is evaluated according to the signal-to-noise ratio and image clarity quantitative indicators of the first infrared temperature map, and a second infrared temperature map is obtained by processing with a wavelet transform denoising algorithm; the contrast of the second infrared temperature map is enhanced by a Laplace operator, and a region is segmented by a region growing algorithm and an edge detection algorithm to form a region composed of points whose temperature is higher than a preset threshold value of the background temperature to obtain a heat source candidate region map; the area size, contour shape, edge curvature and temperature distribution characteristic parameters of each candidate region are extracted from the heat source candidate region map, and a target type mark and a temperature distribution characteristic parameter infrared target recognition result map are generated according to the animal and human characteristic parameters in a pre-established heat source characteristic database.

[0018] Specifically, an infrared thermal imager is used to obtain an infrared image in front of the vehicle, a temperature correction curve is established by calibrating the temperature of the black body radiation source and the actual temperature, and the infrared image is corrected for brightness and temperature to generate a first infrared temperature map. The first infrared temperature map is evaluated for image quality, and the image quality level is evaluated by signal-to-noise ratio calculation and image clarity quantitative indicators. When the quality level reaches a preset quality threshold, a wavelet transform denoising algorithm is used to obtain a second infrared temperature map. The contrast of the second infrared temperature map is enhanced by a Laplace operator, and a region consisting of points with a temperature higher than a preset threshold of the background temperature is segmented by a region growing algorithm and an edge detection algorithm to obtain a heat source candidate region map. The area size, contour shape, edge curvature and temperature distribution feature parameters of each candidate region are extracted from the heat source candidate region map, and the candidate region category determination result is obtained based on the support vector machine classification algorithm according to the animal and human feature parameters in the pre-established heat source feature database. For the candidate region category determination result, if it is determined to be an animal, it is marked as an animal target region, and if it is determined to be a human, it is marked as a human target region, and a target type marking map is generated. According to the temperature distribution gradient and regional edge features of the animal target area or the human target area in the target type marking image, the target posture feature parameters are identified through the skeleton extraction algorithm, and the target temperature distribution feature vector is established. The target temperature distribution feature vector is superimposed on the target type marking image to generate an infrared target recognition result image containing the target position, type, posture information and temperature distribution feature parameters. The infrared thermal imager obtains the temperature distribution information of the target by detecting the infrared energy radiated by the target. For homeothermic organisms such as animals and humans, their surface temperature is usually higher than the ambient temperature, which appears as an obvious heat source area in the infrared image. The original image signal collected by the infrared thermal imager needs to be temperature corrected. The radiation intensity measured at different temperature points by a calibrated black body is used to establish a correction curve with the actual temperature value. The relationship between the black body radiation intensity and temperature conforms to the Stefan-Boltzmann law. When the temperature is in the range of 283K to 313K, a correction point is recorded every 5K, and the temperature correction curve is obtained by least squares fitting. The signal-to-noise ratio and clarity indicators are used to evaluate the quality of infrared images. The signal-to-noise ratio SNR calculation formula is:

[0019] SNR = 20lg (As / An), where As is the image signal amplitude and An is the noise amplitude. When the signal-to-noise ratio is greater than 35dB, it indicates that the image quality is good. Image clarity is evaluated by calculating the variance and gradient of the image. The larger the variance, the higher the image contrast, and the larger the gradient value, the clearer the edge. The image variance threshold is set to 225 and the gradient threshold is set to 0.45. The heat source region segmentation adopts an improved regional growing algorithm, selects points with a temperature higher than the ambient temperature by more than 8K as seed points, and expands to the surrounding in an 8-neighborhood manner. If the temperature difference between adjacent pixels is less than 1.5K, they are merged into the current region. Edge detection uses the Laplace operator for image enhancement to highlight the edge contour features of the target. Target feature extraction includes shape features such as area, perimeter, and circularity, as well as temperature distribution features. The calculation formula of circularity C is C = 4πA / L2, where A is the area of ​​the region and L is the perimeter. The circularity of human targets is usually between 0.6 and 0.8, while the circularity of animal targets is lower due to the presence of limbs, about 0.4 to 0.6. The temperature distribution characteristics are characterized by calculating the mean, standard deviation and kurtosis of the temperature in the area. The temperature distribution of the human body is relatively uniform, with a standard deviation of less than 2K, while the temperature distribution of animals is more uneven due to their fur. The target posture recognition uses a skeleton extraction algorithm. The target area is extracted as a skeleton line with a single pixel width through a refinement algorithm, and the main direction and branch point position of the skeleton line are calculated. When the human body is standing, it presents an obvious vertical main skeleton line, while animals present a horizontal main skeleton line and multiple branch points. The temperature distribution feature vector contains parameters such as the average temperature of the area, the maximum temperature, the standard deviation of the temperature, and the kurtosis of the temperature distribution. It is combined with the target position coordinates and the posture feature parameters to form a target feature description.

[0020] Step S102 , for the identified animal or human heat source target, track its movement trajectory in real time, obtain movement trajectory data within a predetermined time, analyze the trajectory data through a trajectory clustering algorithm, and obtain the movement trajectory pattern of the animal or human.

[0021] A Kalman filter operator is used to track the target area in the heat source target recognition result map in real time, and a first motion trajectory coordinate sequence is generated according to the centroid coordinates of the target area; according to the first motion trajectory coordinate sequence, the points exceeding the threshold of the preset coordinate range are supplemented by cubic spline interpolation to obtain a second motion trajectory coordinate sequence; for the second motion trajectory coordinate sequence, a Gaussian filter is used to eliminate the coordinate jitter error, and the displacement distance and direction angle between the trajectory points are calculated to obtain a third motion trajectory coordinate sequence; according to the third motion trajectory coordinate sequence, the inflection point coordinates and the trajectory segmentation points are extracted by a density clustering algorithm, and the trajectory feature vector is calculated to obtain the target motion trajectory feature pattern recognition result.

[0022] Specifically, according to the target type marked area in the heat source target recognition result map, the Kalman filter operator is used to obtain the centroid coordinates of the human target or animal target area in real time, and the target centroid position coordinate point data is recorded at a sampling interval of 200 milliseconds to generate a first motion trajectory coordinate sequence. The first motion trajectory coordinate sequence is subjected to outlier detection, and the points exceeding the preset coordinate range threshold are removed. The missing trajectory points are supplemented by cubic spline interpolation to generate a second motion trajectory coordinate sequence. For the second motion trajectory coordinate sequence, a Gaussian filter is used to remove the coordinate jitter error, and the displacement distance and direction angle between adjacent trajectory points are calculated to generate a third motion trajectory coordinate sequence. According to the third motion trajectory coordinate sequence, the trajectory curvature value and angle change are calculated, and the trajectory points are clustered based on the density clustering algorithm, the inflection point coordinates and the trajectory segmentation points are extracted, and the trajectory feature segmentation data is generated. For the trajectory feature segmentation data, the average speed, acceleration, curvature change rate and motion direction change rate of each trajectory segment are calculated to construct a trajectory feature vector. According to the trajectory feature vector, the trajectory feature data is trained by a support vector machine classifier to establish an animal and human motion trajectory model. The animal and human motion trajectory model is used to match the newly acquired trajectory to obtain the target motion trajectory feature pattern recognition result. In moving target tracking, the Kalman filter operator estimates the target position through two stages: prediction and update. The prediction stage predicts the position at the next moment based on the target motion state equation, and the update stage corrects the prediction result based on the actual observation value. The state equation contains the position coordinates and velocity components, and the observation equation reflects the relationship between the sensor measurement value and the actual state. In practical applications, for a uniformly moving target, the state transfer matrix is ​​set to

[0023] [[1,0,dt,0],[0,1,0,dt],[0,0,1,0],[0,0,0,1]], where dt is the sampling time interval of 200 milliseconds. In trajectory data preprocessing, the 3σ criterion is used for outlier detection to calculate the mean and standard deviation of trajectory point coordinates. If the coordinates of a point deviate from the mean by more than 3 times the standard deviation, it is determined to be an outlier. For missing trajectory points, the cubic spline interpolation method is used to supplement them. The interpolation function meets the requirements of position and first-order derivative continuity at the node. The Gaussian filter uses a 5×5 filter kernel with a standard deviation of 1.5 to eliminate trajectory jitter. In the process of trajectory feature extraction, the displacement distance between adjacent trajectory points is calculated by the Euclidean distance, and the direction angle is calculated by the inverse tangent function. The curvature is calculated using the three-point method. For the three consecutive points P1(x1,y1), P2(x2,y2), and P3(x3,y3) in the trajectory point sequence, the area S of the triangle formed by the three points is calculated by the determinant, and the curvature K=4S / (d12×d23×d31), where dij represents the distance between point i and point j. When the density clustering algorithm groups the trajectory points, the neighborhood radius is set to 50 pixels and the minimum point threshold is 8. The density value of each trajectory point is calculated, and the point with a density value greater than the threshold is used as the cluster center. The inflection point extraction is based on the change in direction angle. When the change in direction angle of three consecutive trajectory points exceeds 45 degrees, the middle point is determined as the inflection point. The trajectory feature vector contains 5 dimensions: average speed, maximum acceleration, mean curvature, mean direction angle change rate, and number of trajectory segments. The trajectory of human motion usually shows the characteristics of relatively stable speed and gentle curvature change. Typical values ​​are: average speed 1.2 m / s, maximum acceleration 0.8 m / s2, curvature mean 0.015, direction angle change rate mean 12 degrees / s, and trajectory segment number 2 to 3 segments. Animal motion trajectories show the characteristics of large speed changes and obvious sharp turns. Typical values ​​are: average speed 2.5 m / s, maximum acceleration 2.2 m / s2, curvature mean 0.045, direction angle change rate mean 35 degrees / s, and trajectory segment number 4 to 6 segments. The support vector machine uses the radial basis kernel function, the penalty factor is 10, and the number of training samples is not less than 1000 groups.

[0024] Step S103, predicting the movement trajectory of the animal or person within a predetermined time according to the movement trajectory pattern, so as to obtain the predicted movement position and speed information, and transmitting the predicted information to the acoustic control unit.

[0025] A long short-term memory neural network is used to train the motion trajectory feature data of animals and humans, wherein the input features include a historical trajectory point coordinate sequence and a speed sequence, to obtain trajectory prediction training data; the trajectory prediction training data is processed through a sliding window, the differential value of the position coordinate point sequence in the window is calculated, and the target motion acceleration sequence is obtained; the target motion acceleration sequence is calculated using a numerical integration method to obtain a position prediction coordinate sequence; the position prediction coordinate sequence is smoothed, the target motion direction angle value and the motion speed value are calculated, and the target motion trajectory prediction curve is obtained through a neural network prediction algorithm.

[0026] Specifically, a long short-term memory neural network is used to train the animal and human motion trajectory feature data, the input features include a historical trajectory point coordinate sequence and a speed sequence, the network structure includes an input layer, a hidden layer and an output layer, the number of neurons in the hidden layer is twice the length of the input sequence, and trajectory prediction training data is generated. The trajectory prediction training data is processed by a sliding window, the window length is the forward five-second trajectory data, the sliding step length is one second, and the differential value of the position coordinate point sequence in the window is calculated to obtain the target motion acceleration sequence. According to the target motion acceleration sequence, the target motion trajectory within the next three seconds is calculated by a numerical integration method to obtain a position prediction coordinate sequence. The position prediction coordinate sequence is smoothed, the target motion direction angle value and the motion speed value are calculated, and the target motion trajectory prediction curve is obtained by a neural network prediction algorithm. According to the target motion trajectory prediction curve, the target position coordinate value, the motion speed value, and the motion direction angle value are extracted to generate a target prediction motion data packet. The target prediction motion data packet is encoded by a serial communication protocol, and the prediction data is sent to the acoustic control unit through a serial data interface. The long short-term memory neural network uses a gated structure to model the temporal information, which includes three components: a forget gate, an input gate, and an output gate. The input feature vector contains the position coordinate sequence (x, y) and the velocity sequence (vx, vy) of the target within 5 seconds of history, sampled every 200 milliseconds, with a total of 25 time points. The network hidden layer uses 50 neurons, and the activation function uses the hyperbolic tangent function. It is trained by the stochastic gradient descent algorithm, and the learning rate is set to 0.01. In the sliding window processing, the window length is selected to be 5 seconds corresponding to the target attention duration, and the window slides forward 1 second each time to ensure the continuity of the prediction. The position coordinate sequence is differentially operated within the window, and the acceleration calculation formula between two adjacent points is a = (v2-v1) / dt, where v2 and v1 are the velocities at adjacent moments, and dt is the sampling time interval of 0.2 seconds. For human motion, the typical acceleration value is in the range of -0.8 to 0.8 m / s2, while the acceleration of animal motion can reach -2.2 to 2.2 m / s2. The improved Euler method is used for numerical integration, and the integration step is 0.2 seconds to predict the motion trajectory in the next 3 seconds. The integral calculation formula is v(t+dt)=v(t)+a(t)×dt, x(t+dt)=x(t)+v(t)×dt+0.5×a(t)×dt2, where v(t) and x(t) are the velocity and position at time t, respectively, a(t) is the acceleration at time t, dt is the time step, and x(t+dt) is the new position at time t+dt, that is, the position of the next time step. For uniformly accelerated motion, the predicted trajectory is parabolic, while variable accelerated motion presents more complex curve characteristics. The trajectory prediction curve is smoothed using cubic spline interpolation to ensure the continuity of the position, velocity and acceleration of the curve at the nodes.The motion direction angle is calculated by the position difference between adjacent prediction points, angle θ = arctan((y2-y1) / (x2-x1)). For human motion, the direction angle change rate is usually less than 30 degrees / second, while the direction change rate of animal motion can reach 90 degrees / second. The motion speed value is calculated by displacement and time, v = √((x2-x1)2+(y2-y1)2) / dt. The target prediction motion data packet is encapsulated in a fixed format, including 32 bytes of header information and 96 bytes of data segment. The header information includes the data packet identifier, timestamp, and checksum, and the data segment contains the position coordinates (x, y), speed (vx, vy), and direction angle θ of 15 prediction points. The data is transmitted at a baud rate of 115200 through a serial communication protocol, and the transmission time of each data packet is about 10 milliseconds, which meets the real-time requirements. The sampling period of the prediction data matches the transmission delay to ensure the continuity and timeliness of the data.

[0027] In step S104, the acoustic control unit dynamically adjusts the sound pressure level and frequency of the speaker according to the received predicted movement information and target category to generate an acoustic warning method that matches the predicted movement position, speed and target type.

[0028] An adaptive neural network is used to calculate the distance from the loudspeaker to the predicted target position, and the sound pressure attenuation coefficient is obtained according to the predicted position distance and the sound wave propagation attenuation formula; the sound pressure modulation amplitude is calculated according to the sound pressure attenuation coefficient and the target speed value, and an envelope extraction algorithm is used to generate a sound pressure level parameter sequence; a reference sound waveform corresponding to the target type is extracted from a sound feature database, and the spectrum distribution characteristics are extracted according to the reference sound waveform to obtain the sound spectrum parameters; a sinusoidal carrier is used for phase modulation according to the sound spectrum parameters and the sound pressure level parameter sequence, and a modulated sound signal is generated by a digital frequency synthesizer, and the modulation depth is proportional to the predicted target speed.

[0029] Specifically, according to the target type tag and the target position coordinate value in the predicted mobile data packet, an adaptive neural network is used to calculate the distance from the speaker to the target predicted position, and the sound pressure attenuation coefficient is calculated according to the sound wave propagation attenuation formula to generate a first sound pressure level parameter sequence. The first sound pressure level parameter sequence is normalized, the sound pressure modulation amplitude is calculated according to the target speed value and the displacement distance, and the envelope extraction algorithm is used to generate a second sound pressure level parameter sequence. The corresponding target type reference sound waveform is extracted from the sound feature database, the sound feature database contains animal sound waveforms and human voice warning waveforms, and the reference waveform spectrum distribution characteristics are extracted to obtain the sound spectrum parameters. According to the sound spectrum parameters and the second sound pressure level parameter sequence, a bandpass filter group is used to adjust the gain of each frequency band, and the target sound waveform is generated by the Fourier synthesis algorithm. The target sound waveform is phase modulated by a sine carrier, and the modulation depth is proportional to the target predicted speed, and a modulated sound signal is generated by a digital frequency synthesizer. The modulated sound signal is converted into digital and analog, and the gain control amount is calculated according to the signal amplitude change rate in adjacent control cycles, and the speaker sound unit is driven by a power amplifier to generate an acoustic signal. When sound waves propagate in the air, they follow the distance attenuation law. The sound pressure level attenuates as the distance increases. The attenuation calculation formula is ΔL = 20lg (r2 / r1), where r2 is the predicted position distance and r1 is the reference distance of 1 meter. When the target is 20 meters away from the speaker, the sound pressure attenuation is 26 decibels, and the attenuation reaches 34 decibels at a distance of 50 meters. The sound pressure compensation value is calculated in real time through an adaptive neural network to keep the target receiving sound pressure level within the preset range. The sound pressure modulation amplitude is related to the target movement speed. The envelope extraction algorithm is used to calculate the modulation depth. The modulation depth calculation formula is m = k × v, where k is the proportional coefficient 0.2 seconds / meter, and v is the target speed in meters / second. For walking people, the speed is about 1.2 meters / second and the modulation depth is 0.24; for running animals, the speed can reach 2.5 meters / second and the modulation depth is 0.5. The sound pressure level parameter sequence is normalized to map the amplitude to between 0 and 1. Different types of reference sound waveforms are stored in the sound feature database, with a sampling rate of 44100 Hz and a quantization accuracy of 16 bits. The spectrum of the human voice warning waveform is mainly concentrated in the range of 500 to 2000 Hz, with a strong intermediate frequency component; the animal sound waveform contains a wider frequency band, which can reach 200 to 8000 Hz. The waveform spectrum characteristics are extracted by Fourier transform, and the energy distribution of each frequency band is calculated. The bandpass filter group contains 8 sub-band filters, with center frequencies of 250, 500, 1000, 2000, 4000, and 8000 Hz, respectively, and the bandwidth is 0.2 times the center frequency. The gain of each filter is adjusted according to the spectrum parameters, and the gain range is -12 to 12 decibels. Fourier synthesis uses inverse Fourier transform to convert the adjusted spectrum into a time domain waveform. Phase modulation uses a sinusoidal carrier, and the carrier frequency is selected to be 20 kHz to avoid the human ear's hearing range.The phase offset of the modulation signal φ(t) = β×s in(2πfmt), where β is the modulation index, which is related to the target speed, and fm is the modulation frequency, which is 2 Hz. The digital frequency synthesizer uses a direct digital synthesis method to generate a modulated carrier through a phase accumulator. The power amplifier uses a class D amplifier circuit with an operating frequency of 100 kHz and a gain dynamic range of 40 dB. The adjacent control cycle is 20 milliseconds, and the gain is dynamically adjusted according to the signal amplitude change rate. When the change rate exceeds the preset threshold, a soft start method is used for smooth transition to avoid sound distortion. The speaker sound unit adopts a dual voice coil structure with a frequency response range of 20 to 20,000 Hz, a rated power of 50 watts, and a sound pressure sensitivity of 92 dB.

[0030] In step S105, if the target recognition result is a pedestrian, the acoustic control unit inputs the target type, predicted trajectory, and speed parameters into the pre-trained decision neural network, learns based on historical data, generates a warning strategy for pedestrians, and obtains the real-time frequency and volume adjustment curve of the prompt sound. The acoustic control unit generates a dynamic prompt sound signal based on this, which changes continuously as the target approaches to effectively remind pedestrians.

[0031] According to the target recognition result, historical trajectory parameters are extracted from a preset pedestrian trajectory database, and the historical trajectory parameters are processed by a deep neural network to obtain a pedestrian trajectory feature vector; for the pedestrian trajectory feature vector, the historical trajectory parameters are trained by a recursive neural network, and trajectory prediction data is generated by a time series expansion method; for the target distance, speed, and direction parameters in the trajectory prediction data, the trajectory jitter is eliminated by an adaptive filter, and a pedestrian trajectory smooth curve is obtained by a Kalman filter algorithm; for the pedestrian trajectory smooth curve, a reinforcement learning algorithm is used to train the warning data in the historical warning database, and the sound frequency modulation parameters and volume modulation parameters are generated by an acoustic warning decider, and the digital power amplifier drives the speaker array to generate an acoustic signal.

[0032] Specifically, according to the pedestrian target recognition result and the target type tag, historical trajectory parameters are extracted from a preset pedestrian trajectory database, and the predicted trajectory data is processed by a deep neural network to generate a pedestrian trajectory feature vector. The pedestrian trajectory feature vector is standardized, the trajectory features are time-series expanded by a stacked decoder, and the historical trajectory data is trained by a recursive neural network to generate pedestrian trajectory prediction data. According to the target distance, speed, and direction parameters in the pedestrian trajectory prediction data, an adaptive filter is used to eliminate trajectory jitter, and a pedestrian trajectory smooth curve is calculated by a Kalman filter. For the pedestrian trajectory smooth curve, historical warning data is trained based on a reinforcement learning algorithm, an acoustic warning decision maker is constructed, and a sound frequency modulation parameter is generated. According to the sound frequency modulation parameter, a reference prompt audio is synthesized by a digital waveform generator, and an envelope detector is used to extract an audio energy curve to generate a volume modulation parameter. The volume modulation parameter is distance compensated, the spectrum distribution is adjusted by a multi-bandpass filter group, and a digital power amplifier is used to drive a speaker array to generate a dynamically changing acoustic signal. The pedestrian trajectory feature vector contains time series data of four dimensions: position, speed, acceleration, and direction. Each dimension is represented by a 32-bit floating point number. The deep neural network adopts an encoder-decoder structure. The encoder contains three convolutional layers with convolution kernel sizes of 3×3, 5×5, and 7×7, respectively. Each layer is followed by a ReLU activation function and a batch normalization layer. The decoder adopts a deconvolution structure and outputs a 128-dimensional feature vector. The stacked decoder performs temporal expansion of the trajectory features, with a time window length of 5 seconds and a step size of 0.2 seconds. The recurrent neural network adopts a bidirectional long short-term memory network structure with 256 hidden layer neurons. The initial bias values ​​of the input gate and the forget gate are set to 1.0 to avoid gradient disappearance. The training data contains 1,000 sets of pedestrian trajectory data in different scenarios, each set containing feature data at 25 time points. The adaptive filter adopts the minimum mean square error criterion, the filter order is 8, and the step factor μ=0.01. The Kalman filter state vector contains position and velocity components, and the state transfer matrix F = [[1, dt], [0, 1]], where dt is the sampling interval of 0.2 seconds. The diagonal elements of the measurement noise covariance matrix R are set to 0.01, and the diagonal elements of the process noise covariance matrix Q are 0.001. The reinforcement learning algorithm uses a deep Q network. The state space contains three dimensions: target distance, speed, and approach angle, and the action space contains two dimensions: sound frequency and volume. The reward function design considers the warning effect and sound interference. The warning effect is inversely proportional to the target distance, and the interference is proportional to the volume. The network contains 3 fully connected layers, with 128, 64, and 32 neurons respectively, and the learning rate is set to 0.001. The digital waveform generator uses a direct digital synthesis method, with a clock frequency of 100MHz, a phase accumulator bit width of 32 bits, and a frequency resolution of 0.023Hz.The reference prompt audio is synthesized by sine wave superposition, and the base frequency is set in the range of 1kHz to 2kHz. It contains 3 harmonic components, and the amplitude ratio of each harmonic is 1:0.5:0.25. The envelope detection adopts the peak detection method, and the detection time constant is 20ms. The bandpass filter group contains 6 sub-bands, with center frequencies of 500Hz, 1kHz, 2kHz, 4kHz, and 8kHz, and the bandwidth is 0.3 times the center frequency. The filter adopts Butterworth structure with an order of 4. Distance compensation adopts the inverse square law, and the compensation amount ΔL=20lg(r / r0), where r is the target distance and r0 is the reference distance of 1 meter. The power amplifier adopts a Class D structure, with a switching frequency of 200kHz, an output power of 50W, and a total harmonic distortion of less than 0.1%. The speaker array adopts a linear arrangement with a unit spacing of 0.1 meters. The directivity control of the sound beam is achieved by adjusting the phase of each unit.

[0033] If the predicted pedestrian movement trajectory intersects with the vehicle's travel direction and the intersection distance is less than the safety threshold given by the network, the prompt sound frequency and volume will be increased to effectively remind pedestrians to pay attention to avoid; if the predicted pedestrian movement trajectory is parallel to the vehicle's travel direction and the distance is less than the safety threshold given by the network, a prompt sound with moderate volume and moderately increased frequency will be generated to remind pedestrians that the vehicle is approaching.

[0034] According to the vehicle driving direction vector and the predicted pedestrian trajectory vector, the spatial analytic geometry method is used to calculate the coordinates of the intersection of the trajectory extension lines to obtain the first trajectory risk parameter and the second trajectory risk parameter; for the first trajectory risk parameter and the second trajectory risk parameter, the historical safety threshold data is processed through a deep neural network to obtain a safety risk level determination result; if the safety risk level determination result shows that the trajectory intersection distance is less than the safety threshold, the sound frequency growth curve is calculated to generate the sound modulation parameters; for the sound modulation parameters, the frequency spectrum distribution is adjusted through a Butterworth bandpass filter group, and a digital power amplifier is used to drive the speaker array to generate an acoustic prompt signal.

[0035] Specifically, according to the vehicle driving direction vector and the predicted pedestrian trajectory vector, the spatial analytical geometry method is used to calculate the coordinates of the intersection of the trajectory extension line to generate a first trajectory risk parameter. The first trajectory risk parameter is numerically analyzed to calculate the angle between the vehicle driving direction and the pedestrian movement direction to generate a second trajectory risk parameter. According to the first trajectory risk parameter and the second trajectory risk parameter, the historical safety threshold data is processed by a deep neural network to generate a safety risk level determination result. If the safety risk level determination result shows that the trajectory intersection distance is less than the safety threshold, the exponential function y=2^x is used to calculate the sound frequency growth curve and the volume amplification factor to generate a first sound modulation parameter. If the safety risk level determination result shows that the trajectory direction angle is less than the angle threshold, the linear function y=kx+b is used to calculate the sound frequency growth curve and the volume amplification factor to generate a second sound modulation parameter. According to the first sound modulation parameter or the second sound modulation parameter, a digital synthesizer is used to generate a reference prompt audio, and the waveform frequency is dynamically adjusted by a sine wave modulator to generate a modulated audio signal. For the modulated audio signal, the spectrum distribution is adjusted by a Butterworth bandpass filter group, and a digital power amplifier is used to drive the speaker array to generate an acoustic prompt signal. When calculating the intersection of trajectories by spatial analytical geometry, the vehicle direction vector is represented as v1(x1,y1), and the pedestrian trajectory vector is represented as v2(x2,y2). The equations of the two trajectory lines are y=k1x+b1 for l1 and y=k2x+b2 for l2, where the slopes are k1=y1 / x1 and k2=y2 / x2. Solving the equations, we get the coordinates of the intersection (x,y), x=(b2-b1) / (k1-k2), y=k1x+b1. When k1=k2, it means that the two trajectories are parallel. In the trajectory risk parameter calculation, the direction angle

[0036] θ=arccos((v1·v2) / (|v1|·|v2|)), when θ<30 degrees, it is judged as parallel, and when θ>60 degrees, it is judged as intersection. For the intersection, the distance d from the intersection to the current position of the vehicle is calculated. The safety threshold is dynamically adjusted with the speed. The baseline value is 20 meters, and the threshold is increased by 5 meters for every 10 km / h increase in speed. The deep neural network adopts a feedforward structure. The input layer contains four parameters: trajectory intersection distance, direction angle, vehicle speed, and pedestrian speed. The hidden layer adopts a 3-layer structure, with 64, 32, and 16 neurons respectively, and the activation function adopts ReLU. The output layer outputs the risk level probability distribution through the Softmax function, which is divided into three levels: low, medium, and high. When it is judged as a high-risk intersection, the frequency growth curve uses the exponential function y=2^x, x is the normalized distance parameter, the distance from the safety threshold to 0 corresponds to x from 0 to 1, and the frequency increases from 1000 Hz to 4000 Hz. The volume growth also adopts an exponential law, increasing from 60 decibels to 90 decibels. For the parallel approach situation, a linear growth function y=kx+b is used, k=-30, b=90, and the volume changes linearly from 60 decibels to 75 decibels from the safety threshold to 5 meters. The digital synthesizer adopts a direct digital synthesis method to generate a reference frequency signal through a phase accumulator, with a clock frequency of 100 MHz and a frequency resolution of 0.023 Hz. Sine wave modulation adopts frequency modulation, and the instantaneous frequency ω(t)=ω0+kf×m(t), where ω0 is the carrier frequency, kf is the frequency modulation sensitivity, and m(t) is the modulation signal. The carrier frequency is selected as 2000 Hz, and the frequency modulation range is ±1000 Hz. The Butterworth filter group contains 3 sub-bands with center frequencies of 1000, 2000, and 4000 Hz, respectively. The bandwidth is 0.5 times the center frequency, and the filter order is 4. The gain control adopts a piecewise linear function, with 0-6 dB in the low frequency band, 3-9 dB in the mid-frequency band, and 6-12 dB in the high frequency band. The power amplifier adopts a Class D structure, with a switching frequency of 200 kHz, a maximum output power of 50 watts, and a total harmonic distortion of less than 0.1%. The speaker array adopts a 4-unit linear arrangement with a unit spacing of 0.2 meters. The sound beam directivity control is achieved by adjusting the phase of each unit.

[0037] In step S106, if the target recognition result is an animal or a vehicle, the acoustic control unit inputs the target parameters into the pre-trained reinforcement learning model to obtain the optimal warning time point, volume step adjustment amplitude and frequency. The acoustic control unit generates a horn signal in real time based on this, and adaptively adjusts it according to the target reaction until the safe avoidance goal is achieved.

[0038] A reinforcement learning neural network is trained based on historical avoidance data, and the reinforcement learning neural network generates first target feature data from three dimensions: target distance, speed, and movement direction; a sliding time window is used to process the target trajectory data, and the sliding time window generates second target feature data by calculating target position, speed, and acceleration parameters; a reward function is established based on the first target feature data, and the reward function processes the second target feature data through a Bayesian decision maker to obtain an optimal warning timing; acoustic control parameters are calculated for the optimal warning timing, and the acoustic control parameters generate dynamic prompt audio through a digital waveform generator, and the dynamic prompt audio is updated in real time according to the target response feature vector.

[0039] Specifically, according to the target type labeling result and the target prediction trajectory data, the reinforcement learning neural network is used to train the historical avoidance data. The state space includes three dimensions: target distance, speed, and movement direction, and the first target feature data is generated. The target trajectory data is processed through a five-second sliding time window, and the window step is set to one second. The current position, speed, and acceleration parameters of the target are calculated to generate the second target feature data. The second target feature data is standardized, and a reward function is established according to the first target feature data. The Bayesian decision maker is used to generate the optimal warning time. According to the optimal warning time, it is compared with the preset reference acoustic parameter group, and the volume and frequency adjustment amount are calculated using a piecewise function to generate acoustic control parameters. For the acoustic control parameters, a reference prompt audio is generated by a digital waveform generator, and the audio signal is modulated by a step function to generate a dynamic prompt audio. According to the infrared image feature identifier, the target avoidance state is detected in real time, the target displacement increment and direction change are calculated, and the target response feature vector is generated. The target response feature vector is processed by a dynamic programming algorithm, the acoustic control parameters are updated in real time, and the dynamic prompt audio is adaptively adjusted to generate the final warning signal. The reinforcement learning neural network adopts a deep Q learning structure. The state space includes three dimensions: target distance, speed, and direction of movement. The distance range is 0 to 50 meters, the speed range is 0 to 20 meters / second, and the direction angle range is 0 to 360 degrees, which is discrete into 50×20×36 states. The action space includes three dimensions: warning timing, volume, and frequency, and each dimension is discrete into 10 values. The reward function design takes into account the avoidance effect and sound interference. The avoidance effect is related to the displacement direction angle, and the interference is related to the sound pressure level. The specific form is R=k1×cos(θ)-k2×(L-L0), where θ is the angle between the displacement direction and the expected direction, L is the sound pressure level, L0 is the reference sound pressure level, R is the final reward value or score, and k1 and k2 are weight coefficients, which are used to adjust the relative importance of avoidance effect and sound interference in the total reward. The time window uses a length of 5 seconds to analyze the trajectory data, and is updated every 1 second to maintain the continuity and real-time nature of the data. The target motion parameters are calculated within the window, and the position coordinates are fitted using the least squares method to obtain a smooth trajectory curve. The speed is calculated by position difference, and the acceleration is calculated by speed difference. For vehicle targets, the typical acceleration range is -3 to 3 meters per second squared. For animal targets, the acceleration range is larger, up to -5 to 5 meters per second squared. The Bayesian decision maker calculates the optimal warning time based on conditional probability, P(t|s) = P(s|t) × P(t) / P(s), where t is the warning time and s is the current state. The prior probability P(t) is obtained from historical data statistics, and the likelihood function P(s|t) is modeled using Gaussian distribution. P(s) is the total probability of the current state s occurring, which is the weighted average of the probability of s occurring at all possible warning times t.For vehicle targets, the optimal warning distance is proportional to the speed, and the empirical formula is d=v×t0+d0, where t0 is the benchmark reaction time of 2 seconds and d0 is the benchmark safety distance of 10 meters. The preset benchmark acoustic parameter group contains three levels, corresponding to different levels of urgency. The first level volume is 70 decibels and the frequency is 1000 Hz; the second level volume is 80 decibels and the frequency is 2000 Hz; the third level volume is 90 decibels and the frequency is 4000 Hz. The volume modulation uses a piecewise linear function and the frequency modulation uses an exponential function. For animal targets, the sound frequency range is selected between 2000 and 6000 Hz to avoid their auditory sensitive areas. The target avoidance state recognition uses an infrared image processing method to extract the target contour feature point sequence and calculate the displacement vector of the feature points between adjacent frames. The displacement increment is obtained by vector superposition, and the direction change is calculated by the vector angle. When the displacement increment is greater than 0.5 meters and the direction angle is greater than 30 degrees, it is determined that the target begins to avoid. The dynamic programming algorithm adjusts the acoustic parameters in real time according to the avoidance status, and the adjustment step size is proportional to the avoidance degree to avoid parameter oscillation. The adaptive adjustment process continues until the target displacement exceeds the safe distance or the direction deviates from the dangerous area.

[0040] If the predicted animal / vehicle movement trajectory intersects with the vehicle's travel direction and the predicted collision time is less than the safety threshold, a horn signal with the maximum volume is generated to prompt the animal / vehicle to take emergency avoidance measures; if the predicted animal / vehicle movement trajectory is parallel to the vehicle's travel direction and the distance is less than the safety threshold, the volume is gradually increased to guide the animal / vehicle to avoid in advance until both parties are at a safe distance.

[0041] Vector analysis is performed based on the vehicle driving direction vector and the target predicted trajectory vector to obtain the first trajectory feature data of the trajectory intersection coordinates and the intersection angle; for the first trajectory feature data, historical collision data is processed through a feedforward neural network, and the feedforward neural network input layer includes three parameters: trajectory intersection distance, intersection angle, and relative speed, to obtain the collision prediction time; if the collision prediction time is less than the safety time threshold and the trajectory intersection angle is greater than the angle threshold, a maximum sound pressure level control parameter is generated; if the collision prediction time is greater than the safety time threshold and the trajectory intersection angle is less than the angle threshold, a progressive sound pressure level control parameter is generated; an infrared image processor is used to detect the target position coordinate sequence, and avoidance feature data is obtained by calculating the target displacement vector and the direction angle change, and the sound pressure level control parameter is updated in real time through an adaptive filter according to the avoidance feature data.

[0042] Specifically, according to the vehicle driving direction vector and the target predicted trajectory vector, the vector analysis method is used to calculate the trajectory intersection coordinates and the intersection angle to generate the first trajectory feature data. For the first trajectory feature data, the historical collision data is processed by a feedforward neural network, and the input layer contains three parameters: trajectory intersection distance, intersection angle, and relative speed, to generate the collision prediction time. According to the comparison result of the collision prediction time and the safety time threshold, if the prediction time is less than the safety time threshold and the trajectory intersection angle is greater than the angle threshold, a step function is used to generate the maximum sound pressure level control parameter. If the collision prediction time is greater than the safety time threshold and the trajectory intersection angle is less than the angle threshold, a piecewise linear function is used to generate the progressive sound pressure level control parameter. According to the maximum sound pressure level control parameter or the progressive sound pressure level control parameter, a reference acoustic signal is generated by a digital waveform generator, and a dynamic acoustic signal is generated by a volume modulator. An infrared image processor is used to detect the target position coordinate sequence in real time, and the target displacement vector and the direction angle change are calculated to generate the avoidance feature data. According to the avoidance characteristic data, the avoidance degree parameter is calculated by an adaptive filter, the sound pressure level control parameter is updated in real time, and the dynamic acoustic signal is optimized and adjusted. In the vector analysis calculation, the intersection angle between the vehicle driving direction vector v1 (x1, y1) and the target trajectory vector v2 (x2, y2) is calculated by the vector angle formula, θ = arccos ((v1·v2) / (|v1|·|v2|)), v1·v2 is the dot product of v1 and v2, |v1| and |v2| are the modulus lengths of vectors v1 and v2. The coordinates of the trajectory intersection are obtained by solving a system of linear equations. When the equations of the two trajectory lines are y = k1x + b1 and y = k2x + b2, the x coordinate of the intersection is x = (b2-b1) / (k1-k2). For the case of a vehicle speed of 20 m / s and a target speed of 10 m / s, when the intersection angle is 90 degrees, a collision is expected to occur after 5 seconds. The feedforward neural network adopts a three-layer structure. The three neurons in the input layer correspond to the track intersection distance, intersection angle, and relative speed. The hidden layer uses 64 neurons, and the output layer uses one neuron to output the collision prediction time. ReLU is used as the activation function to avoid the gradient vanishing problem. The training data contains 1,000 sets of historical collision records, which are trained using the stochastic gradient descent algorithm. The learning rate is set to 0.01, and the training error is less than 0.1 seconds. The safety time threshold is dynamically set according to the target type, 3 seconds for vehicle targets and 2 seconds for animal targets. The angle threshold is set to 30 degrees. If it is greater than this value, it is judged as an intersecting track, and if it is less than this value, it is judged as a parallel track. The step function directly increases the sound pressure level to the maximum value of 90 decibels when it is judged as a dangerous situation. The piecewise linear function gradually increases the sound pressure level from 60 decibels to 75 decibels according to the distance. The digital waveform generator adopts the direct digital synthesis method to generate a reference frequency signal through a phase accumulator. The clock frequency is 100 MHz and the frequency resolution is 0.023 Hz.The acoustic signal adopts dual-frequency synthesis, and 1000 Hz and 2000 Hz are selected for vehicle targets, and 2000 Hz and 4000 Hz are selected for animal targets. The volume modulation adopts exponential envelope, and the modulation depth is proportional to the urgency of avoidance. The infrared image processing adopts the target tracking algorithm to extract the target centroid coordinate sequence and calculate the displacement vector between adjacent frames. The angle between the displacement direction and the expected avoidance direction is used as the avoidance effect evaluation index. When the angle is less than 45 degrees and the displacement speed is greater than 1 m / s, it is determined that the target begins to avoid effectively. The adaptive filter dynamically adjusts the filter coefficient according to the avoidance degree. The filter order is 8 and the step factor μ=0.01. The avoidance feature data includes three dimensions: displacement speed, direction angle, and acceleration. The measurement noise is eliminated by Kalman filtering. The avoidance degree parameter is calculated by fuzzy rules. The input variable is the normalized value of the three feature dimensions, and the output variable is the sound pressure level adjustment. When the avoidance degree reaches the expected target, the sound pressure level gradually decreases until it returns to the ambient background sound pressure level.

[0043] Step S107, while sending out the acoustic signal, continuously track the actual moving trajectory of the animal or person, and optimize the trajectory prediction method by comparing the deviation between the actual trajectory and the predicted trajectory to achieve the best avoidance effect.

[0044] A target tracker is used to obtain a center of mass coordinate sequence of an animal or human target, and the center of mass coordinate sequence of the target is smoothed by a Kalman filter to obtain first measured trajectory data; the displacement increment and velocity component in the trajectory segment are calculated according to the first measured trajectory data, and the trajectory curve is fitted by a recursive least squares method to obtain second measured trajectory data; the Euclidean distance between the measured trajectory point and the predicted trajectory point is calculated for the second measured trajectory data, and the trajectory prediction error vector is calculated by a weighted average method to obtain trajectory correction parameters; the current position and speed of the target are predicted and calculated according to the trajectory correction parameters, and the horn sound is adjusted by a fuzzy controller by calculating the rate of change of the angle between the target moving trajectory and the vehicle driving direction.

[0045] Specifically, a target tracker is used to obtain a centroid coordinate sequence of an animal or human target in real time, and the trajectory segment is divided according to a fixed time window. The target coordinate sequence is smoothed by a Kalman filter to generate the first measured trajectory data. The first measured trajectory data is feature extracted, the displacement increment and velocity component in the trajectory segment are calculated, and the trajectory curve is fitted by the recursive least squares method to generate the second measured trajectory data. According to the second measured trajectory data, the Euclidean distance between the measured trajectory point and the predicted trajectory point is calculated, and the trajectory prediction error vector is calculated by the weighted average method to generate the trajectory correction parameter. For the trajectory correction parameter, a deep neural network is used for online learning, and the input layer includes three parameters: position error, velocity error, and direction error. The network weight is updated by the back propagation algorithm. According to the updated neural network, the current position and velocity of the target are predicted and calculated to generate the optimized predicted trajectory data. The optimized predicted trajectory data is evaluated in real time, and the change rate of the angle between the target moving trajectory and the vehicle driving direction is calculated to generate the trajectory evaluation parameter. According to the trajectory evaluation parameter, a fuzzy controller is used to generate the sound pressure level adjustment amount and the frequency modulation parameter, and the horn sound is optimized and adjusted in real time. The target tracker uses infrared image processing to obtain the target center of mass coordinates, with a sampling frequency of 50 Hz, a fixed time window length of 5 seconds, and an overlap of 2.5 seconds for adjacent windows. The Kalman filter state vector contains position and velocity components, and the state transfer matrix F = [[1, dt], [0, 1]], where dt is the sampling interval of 0.02 seconds, the diagonal elements of the measurement noise covariance matrix R are set to 0.01, and the diagonal elements of the process noise covariance matrix Q are 0.001. The trajectory prediction error is calculated using the weighted Euclidean distance d, d = √(w1Δx 2 +w2Δy2+w3Δvx 2+w4Δvy2), where Δx and Δy are position errors, Δvx and Δvy are speed x and y errors respectively, and weight coefficients w1=w2=1.0 and w3=w4=0.5. The error vector is calculated by exponential weighted averaging, and α=0.7 is the smoothing coefficient. The deep neural network adopts a three-layer structure, with 3 neurons in the input layer, 64 neurons in the hidden layer, and 2 neurons in the output layer corresponding to the predicted position increment. The activation function adopts ReLU, the loss function is the mean square error, the initial value of the learning rate is 0.01, and it decays with the training rounds. The back propagation adopts the stochastic gradient descent method, and the mini-batch size is 32. The trajectory evaluation calculates the angle θ=arccos((v1·v2) / (|v1|·|v2|)) between the vehicle direction vector and the target moving direction vector, and the angle change rate dθ / dt is calculated by difference. When the change rate is greater than the preset threshold of 0.5 radians / second, it indicates that the target begins to turn to avoid. The input variables of the fuzzy controller are the angle θ and the rate of change dθ / dt, and the output variables are the sound pressure level adjustment ΔL and the frequency modulation depth m. The fuzzy rule uses Mamdani reasoning, the membership function uses triangles, and the rule base contains 25 rules. The sound pressure level adjustment range is -12 to 12 decibels, and the frequency modulation depth range is 0 to 1. The defuzzification uses the centroid method, and the control cycle is 20 milliseconds. For the case of good avoidance effect, the sound pressure level gradually decreases and the modulation depth decreases; for the case of insufficient avoidance, the sound pressure level is maintained or increased, and the modulation depth increases. In practical applications, if the target is a deer moving on the road, the tracker will collect the position coordinates of the deer every 0.02 seconds. In order to reduce the influence of measurement noise, the Kalman filter is used to smooth the original coordinate sequence. For example, when the deer moves from one side of the road to the other side, the trajectory may appear jagged due to measurement errors. After filtering, a smoother trajectory curve can be obtained. Feature extraction is performed on the smoothed trajectory data to calculate the displacement and velocity between adjacent sampling points. Taking the movement of a deer as an example, if the horizontal displacement between two sampling points is 0.5 meters and the vertical displacement is 0.3 meters, then the average speed of the deer during this period can be calculated. By fitting the trajectory curve using the recursive least squares method, a more accurate description of the movement trajectory can be obtained. The measured trajectory is compared with the predicted trajectory to calculate the distance deviation between the trajectory points. Assuming that the predicted trajectory shows that the deer will continue to move in a straight line, but the actual trajectory shows that the deer begins to turn, this deviation will be used to update the prediction model to optimize the prediction accuracy. When the prediction error is found to be large, the neural network will automatically adjust the weight parameters. The trajectory evaluation stage focuses on the relative motion relationship between the target and the vehicle. For example, when the angle between the deer's moving direction and the vehicle's driving direction begins to increase, it indicates that the deer is avoiding. At this time, the angle change rate can be calculated to determine the effectiveness of the avoidance behavior. If the angle change rate reaches 0.5 radians per second, it means that the deer's avoidance action is relatively obvious. According to the trajectory evaluation results, the fuzzy controller will adjust the parameters of the warning sound accordingly.When the deer begins to avoid the animal effectively, the system will gradually reduce the sound pressure level and the frequency modulation depth to make the sound softer. On the contrary, if the deer does not avoid the animal sufficiently, the sound pressure level will be maintained or increased, and the modulation depth will be increased to make the sound more warning. This dynamic adjustment ensures the warning effect while avoiding excessive disturbance.

[0046] The above only lists some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and changes can be made. As long as the improvements and changes are made on the basis of the basic principles of the present invention, they should be regarded as falling within the protection scope of the present invention.

Claims

1. A vehicle automatic sound control method based on infrared recognition, characterized in that: The method comprises: Obtain infrared images within a predetermined range in front of the vehicle, identify heat source targets in the image, and determine whether the heat source targets are animals or humans based on their shape, size, and temperature distribution characteristics; For the identified heat source targets of animals or people, their movement trajectories are tracked in real time, and the movement trajectory data within a predetermined time is obtained. The trajectory data is analyzed through the trajectory clustering algorithm to obtain the movement trajectory pattern of the animals or people; According to the movement trajectory pattern of the animal or person, predict the movement trajectory within the predetermined time to obtain predicted movement position and speed information, and transmit the predicted information to the acoustic control unit; The acoustic control unit dynamically adjusts the sound pressure level and frequency of the horn based on the received prediction information and target category to generate an acoustic warning method that matches the predicted moving position, speed and target type; If the target is identified as a pedestrian, the acoustic control unit inputs the target type, predicted trajectory, and speed parameters into the pre-trained decision neural network, learns based on historical data, generates a warning strategy for pedestrians, and obtains the real-time frequency and volume adjustment curve of the prompt sound. The acoustic control unit generates a dynamic prompt sound signal based on this, which changes continuously as the target approaches; If the target recognition result is an animal or a vehicle, the acoustic control unit inputs the target parameters into the pre-trained reinforcement learning model to obtain the optimal warning time point, volume step adjustment amplitude and frequency. The acoustic control unit generates a horn signal in real time based on the optimal warning time point, volume step adjustment amplitude and frequency, and adaptively adjusts it according to the target's reaction until the safe avoidance goal is achieved.

2. The method according to claim 1, characterized in that The method of acquiring an infrared image within a predetermined range in front of the vehicle, identifying a heat source target in the image, and determining whether the heat source target is an animal or a person according to the shape, size, and temperature distribution characteristics of the heat source target includes: An infrared thermal imager is used to obtain an infrared image in front of the vehicle, a temperature correction curve is established by calibrating the temperature of a black body radiation source and an actual temperature, and brightness correction and temperature correction are performed on the infrared image to obtain a first infrared temperature map; The image quality level is evaluated according to the signal-to-noise ratio and image clarity quantitative index of the first infrared temperature image, and a second infrared temperature image is obtained by processing with a wavelet transform denoising algorithm; For the second infrared temperature map, the contrast is enhanced by the Laplace operator, and the region growing algorithm and the edge detection algorithm are used to segment the regions formed by points whose temperature is higher than the preset threshold value of the background temperature, so as to obtain a heat source candidate region map; The area size, contour shape, edge curvature and temperature distribution characteristic parameters of each candidate area are extracted from the heat source candidate area map, and the target type mark and temperature distribution characteristic parameter infrared target recognition result map are generated according to the animal and human characteristic parameters in the pre-established heat source characteristic database.

3. The method according to claim 1, characterized in that The method of tracking the movement trajectory of the identified animal or human heat source target in real time, obtaining the movement trajectory data within a predetermined time, and analyzing the trajectory data by a trajectory clustering algorithm to obtain a typical movement trajectory pattern of the animal or human includes: A Kalman filter operator is used to track the target area in the heat source target recognition result map in real time, and a first motion trajectory coordinate sequence is generated according to the centroid coordinates of the target area; According to the first motion trajectory coordinate sequence, points exceeding a preset coordinate range threshold are supplemented by cubic spline interpolation to obtain a second motion trajectory coordinate sequence; For the second motion trajectory coordinate sequence, a Gaussian filter is used to eliminate coordinate jitter errors, and the displacement distance and direction angle between trajectory points are calculated to obtain a third motion trajectory coordinate sequence; According to the third motion trajectory coordinate sequence, the inflection point coordinates and trajectory segmentation points are extracted by using a density clustering algorithm, and the trajectory feature vector is calculated to obtain a target motion trajectory feature pattern recognition result.

4. The method according to claim 1, characterized in that The method of predicting the movement trajectory of the animal or person within the predetermined time according to the movement trajectory pattern of the animal or person to obtain the predicted movement position and speed information, and transmitting the predicted information to the acoustic control unit includes: The long short-term memory neural network is used to train the feature data of animal and human motion trajectories, where the input features include the historical trajectory point coordinate sequence and speed sequence, to obtain trajectory prediction training data; Processing the trajectory prediction training data through a sliding window, calculating the difference value of the position coordinate point sequence in the window, and obtaining the target motion acceleration sequence; The target motion acceleration sequence is calculated by using a numerical integration method to obtain a position prediction coordinate sequence; The position prediction coordinate sequence is smoothed, the target motion direction angle value and the motion speed value are calculated, and the target motion trajectory prediction curve is obtained through a neural network prediction algorithm.

5. The method according to claim 1, characterized in that The acoustic control unit dynamically adjusts the sound pressure level and frequency of the horn according to the received prediction information and target category to generate an acoustic warning method that matches the predicted moving position, speed and target type, including: Adaptive neural network is used to calculate the distance from the speaker to the predicted target position, and the sound pressure attenuation coefficient is obtained according to the predicted position distance and the sound wave propagation attenuation formula; Calculating the sound pressure modulation amplitude according to the sound pressure attenuation coefficient and the target speed value, and generating a sound pressure level parameter sequence using an envelope extraction algorithm; Extracting a reference sound waveform corresponding to the target type from a sound feature database, and extracting spectrum distribution features according to the reference sound waveform to obtain sound spectrum parameters; A sinusoidal carrier is used for phase modulation according to the sound spectrum parameter and the sound pressure level parameter sequence, and a modulated sound signal is generated by a digital frequency synthesizer, wherein the modulation depth is proportional to the target predicted speed.

6. The method according to claim 1, characterized in that If the target recognition result is a pedestrian, the acoustic control unit inputs the target type, predicted trajectory, and speed parameters into the pre-trained decision neural network, learns based on historical data, generates a warning strategy for pedestrians, and obtains a real-time frequency and volume adjustment curve of the prompt sound. The acoustic control unit generates a dynamic prompt sound signal based on this, which changes continuously as the target approaches, including: According to the target recognition result, historical trajectory parameters are extracted from a preset pedestrian trajectory database, and the historical trajectory parameters are processed by a deep neural network to obtain a pedestrian trajectory feature vector; For the pedestrian trajectory feature vector, the historical trajectory parameters are trained by a recursive neural network, and trajectory prediction data is generated by a time series expansion method; For the target distance, speed and direction parameters in the trajectory prediction data, trajectory jitter is eliminated by an adaptive filter, and a smooth curve of the pedestrian trajectory is obtained by using a Kalman filter algorithm; For the smooth curve of the pedestrian trajectory, a reinforcement learning algorithm is used to train the warning data in the historical warning database, and the sound frequency modulation parameters and volume modulation parameters are generated by the acoustic warning decision maker, and the digital power amplifier drives the speaker array to generate an acoustic signal; It also includes: if the predicted pedestrian movement trajectory intersects with the vehicle's travel direction, and the intersection distance is less than the safety threshold given by the network, then the prompt sound frequency and volume are increased to effectively remind pedestrians to pay attention to avoid; if the predicted pedestrian movement trajectory is parallel to the vehicle's travel direction, and the distance is less than the safety threshold given by the network, then a prompt sound with moderate volume and moderately increased frequency is generated to remind pedestrians that the vehicle is approaching.

7. The method according to claim 6, characterized in that If the predicted pedestrian movement trajectory intersects with the vehicle's driving direction, and the intersection distance is less than the safety threshold given by the network, the frequency and volume of the prompt sound are increased to effectively remind pedestrians to pay attention to avoid; if the predicted pedestrian movement trajectory is parallel to the vehicle's driving direction, and the distance is less than the safety threshold given by the network, a prompt sound with moderate volume and moderately increased frequency is generated to remind pedestrians that the vehicle is approaching, including: According to the vehicle driving direction vector and the predicted pedestrian trajectory vector, the coordinates of the intersection point of the trajectory extension lines are calculated using a spatial analytical geometry method to obtain a first trajectory risk parameter and a second trajectory risk parameter; For the first trajectory risk parameter and the second trajectory risk parameter, the historical safety threshold data is processed by a deep neural network to obtain a safety risk level determination result; If the safety risk level determination result shows that the trajectory intersection distance is less than the safety threshold, the sound frequency growth curve is calculated to generate sound modulation parameters. Based on the sound modulation parameters, the spectrum distribution is adjusted through a Butterworth bandpass filter group, and a digital power amplifier is used to drive the speaker array to generate an acoustic prompt signal.

8. The method according to claim 1, characterized in that If the target recognition result is an animal or a vehicle, the acoustic control unit inputs the target parameters into the pre-trained reinforcement learning model to obtain the optimal warning time point, volume step adjustment amplitude and frequency. The acoustic control unit generates a horn signal in real time according to the optimal warning time point, volume step adjustment amplitude and frequency, and adaptively adjusts according to the target reaction until the safe avoidance goal is achieved, including: Training a reinforcement learning neural network based on historical avoidance data, wherein the reinforcement learning neural network generates first target feature data from three dimensions: target distance, speed, and movement direction; The target trajectory data is processed using a sliding time window, wherein the sliding time window generates second target feature data by calculating target position, velocity and acceleration parameters; Establishing a reward function according to the first target feature data, wherein the reward function processes the second target feature data through a Bayesian decision maker to obtain an optimal warning time; Calculating acoustic control parameters for the optimal warning timing, the acoustic control parameters generating dynamic prompt audio through a digital waveform generator, and the dynamic prompt audio being updated in real time according to a target response feature vector; It also includes: if the predicted animal / vehicle movement trajectory intersects with the vehicle's travel direction and the predicted collision time is less than a safety threshold, a horn signal with the maximum volume is generated to prompt the animal / vehicle to take emergency avoidance measures; if the predicted animal / vehicle movement trajectory is parallel to the vehicle's travel direction and the distance is less than a safety threshold, the animal / vehicle is guided to avoid in advance by gradually increasing the volume until both parties are at a safe distance.

9. The method according to claim 8, characterized in that If the predicted moving trajectory of the animal / vehicle intersects with the driving direction of the vehicle and the predicted collision time is less than the safety threshold, a horn signal with the maximum volume is generated to prompt the animal / vehicle to take emergency avoidance measures; if the predicted moving trajectory of the animal / vehicle is parallel to the driving direction of the vehicle and the distance is less than the safety threshold, the volume is gradually increased to guide the animal / vehicle to avoid in advance until both parties are at a safe distance, including: Performing vector analysis based on the vehicle driving direction vector and the target predicted trajectory vector to obtain first trajectory feature data of trajectory intersection coordinates and intersection angles; For the first trajectory feature data, the historical collision data is processed by a feedforward neural network, wherein the input layer of the feedforward neural network includes three parameters: trajectory intersection distance, intersection angle, and relative speed, to obtain the collision prediction time; If the collision prediction time is less than the safety time threshold and the trajectory intersection angle is greater than the angle threshold, generating a maximum sound pressure level control parameter; If the collision prediction time is greater than a safety time threshold and the trajectory intersection angle is less than an angle threshold, generating a progressive sound pressure level control parameter; An infrared image processor is used to detect the target position coordinate sequence, and avoidance feature data is obtained by calculating the target displacement vector and the direction angle change. The sound pressure level control parameters are updated in real time through an adaptive filter according to the avoidance feature data.

10. The method according to claim 1, characterized in that The method further comprises: A target tracker is used to obtain a centroid coordinate sequence of an animal or human target, and a Kalman filter is used to smooth the centroid coordinate sequence of the target to obtain first measured trajectory data; Calculating the displacement increment and the velocity component in the trajectory segment according to the first measured trajectory data, and fitting the trajectory curve by a recursive least squares method to obtain the second measured trajectory data; Calculating the Euclidean distance between the measured trajectory point and the predicted trajectory point for the second measured trajectory data, calculating the trajectory prediction error vector by using a weighted average method, and obtaining a trajectory correction parameter; The current position and speed of the target are predicted and calculated according to the trajectory correction parameters, and the horn sound is adjusted by using a fuzzy controller by calculating the rate of change of the angle between the target moving trajectory and the vehicle driving direction.

Images