An image processing method for vehicle-mounted infrared pedestrian detection and tracking

Through median filtering, platform histogram equalization, HOG/HOI feature extraction and particle swarm algorithm optimization support vector machine parameters, combined with scale adaptive core-related filtering tracking algorithm, the tracking of vehicle-mounted infrared pedestrian detection tracking when the target scale changes is solved Difficulties and image quality problems achieve high accuracy and stability, and improve the active safety of the vehicle.

CN112288765BActive Publication Date: 2024-11-22JIANGXI CHISHENG INFRARED OPTICS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011198207.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-10-30
Publication Date
2024-11-22
Estimated Expiration
2040-10-30

AI Technical Summary

Technical Problem

The existing vehicle-mounted infrared pedestrian detection and tracking technology cannot effectively track the target scale changes, and the infrared image has a low signal-to-noise ratio and poor contrast, which affects the detection accuracy.

Method used

Median filtering is used to process infrared images, combined with platform histogram equalization and gamma transformation for image enhancement, pedestrian features are extracted using HOG and HOI features, and the support vector machine parameters are optimized through improved particle swarm algorithms, and finally a scale adaptive kernel is used. The relevant filter tracking algorithm performs automatic tracking.

Benefits of technology

It improves the accuracy and tracking stability of on-board infrared pedestrian detection, enhances the real-time and generalization capabilities of detection, and significantly improves the active safety of the vehicle.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112288765B_ABST
    Figure CN112288765B_ABST
Patent Text Reader

Abstract

The present invention discloses an image processing method for on-board infrared pedestrian detection and tracking, including the steps of: 1. filtering the on-board infrared image by median filtering; 2. enhancing the filtered on-board infrared image by platform histogram equalization and gamma transformation; 3. extracting pedestrian features from the enhanced on-board infrared image by HOG features and HOI features, and obtaining edge information and brightness information of pedestrians in the on-board infrared image; 4. optimizing support vector machine parameters by improved particle swarm algorithm for classification, and obtaining the specific position of pedestrians in the on-board infrared image; 5. automatically tracking pedestrians in the on-board infrared image by using scale-adaptive kernel correlation filter tracking algorithm. The method of the present invention has simple steps and is easy to implement. It can be applied in automobile assisted driving systems to achieve accurate detection and stable tracking of on-board infrared pedestrian targets, improve vehicle active safety, have significant effects, and are easy to promote.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to an image processing method for vehicle-mounted infrared pedestrian detection and tracking. Background Art

[0002] With the rapid development of economy and the rapid popularization of automobiles, Automotive Driver Assistance Systems (ADAS) have been widely used. Pedestrian detection and tracking is one of the research directions of ADAS.

[0003] With the development of infrared technology, infrared pedestrian detection has been widely used. Many researchers at home and abroad have conducted research on this topic to varying degrees, and have achieved many results. There are three main methods for pedestrian detection in infrared images: infrared pedestrian detection methods based on region of interest (ROI) segmentation, infrared pedestrian detection methods based on template matching, and infrared pedestrian detection methods based on statistical classification. In the infrared pedestrian detection method based on region of interest segmentation, the acquisition of ROI is a rough classification of pedestrians. Replacing the original image with ROI containing key image features can reduce the influence of a large number of useless backgrounds and noise points, reduce the amount of calculation, improve the real-time performance of the algorithm, and provide a prerequisite for more efficient and accurate detection. However, this type of method also has a major problem, that is, if pedestrians are missed in the process of extracting ROI, the missed pedestrians will not be detected, that is, this method requires a lower missed detection rate to ensure the accuracy of detection. The infrared pedestrian detection method based on template matching needs to design enough templates to achieve a better detection effect. However, the postures of pedestrians are ever-changing, and the types of templates are inexhaustible, so it is impossible to construct a suitable infrared pedestrian target. The infrared pedestrian detection method based on statistical classification mainly extracts the features of positive and negative pedestrian samples in infrared images, trains the classifier, and then traverses the entire image through a sliding window, and classifies each window with the trained classifier, and finally obtains the pedestrian detection result. However, in infrared images, the image quality is fuzzy, and under the premise of ensuring the detection accuracy, there are high requirements for its feature extraction method.

[0004] Target tracking is used to identify moving targets in each consecutive frame of an image and continuously update its position information. In recent years, a large number of target tracking algorithms have emerged. According to their working principles, they are mainly divided into two categories: generative models and discriminative models. At present, discriminative tracking algorithms have become the mainstream direction in the field of tracking. The latest development is the correlation filter tracking algorithm. As far as vehicle-mounted infrared pedestrian tracking is concerned, many of them are improvements on the tracking algorithms in the visible light field. However, the kernel correlation filter algorithm has the problem of being unable to track when the target scale changes. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an image processing method for vehicle-mounted infrared pedestrian detection and tracking in view of the deficiencies in the above-mentioned prior art. The method has simple steps and is easy to implement. It can be applied in automobile assisted driving systems to achieve accurate detection and stable tracking of vehicle-mounted infrared pedestrian targets, thereby improving the active safety of the vehicle. The method has significant effects and is easy to promote.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: an image processing method for vehicle-mounted infrared pedestrian detection and tracking, comprising the following steps:

[0007] Step 1: Use median filtering to filter the vehicle-mounted infrared image;

[0008] Step 2: Use platform histogram equalization and gamma transform to enhance the filtered vehicle-mounted infrared image;

[0009] Step 3: Use HOG features and HOI features to extract pedestrian features from the enhanced vehicle-mounted infrared image to obtain edge information and brightness information of pedestrians in the vehicle-mounted infrared image;

[0010] Step 4: Use the improved particle swarm algorithm to optimize the support vector machine parameters for classification and obtain the specific position of the pedestrian in the vehicle-mounted infrared image;

[0011] Step 5: Use the scale-adaptive kernel correlation filter tracking algorithm to automatically track pedestrians in the vehicle-mounted infrared image.

[0012] In the above-mentioned image processing method for vehicle-mounted infrared pedestrian detection and tracking, the specific process of platform histogram equalization and gamma transformation in step 2 includes:

[0013] Step 201: Determine the threshold value T by taking the average value

[0014]

[0015] Among them, f(x,y) is the filtered vehicle-mounted infrared image, M and N are the sizes of the image;

[0016] Step 202: When the value of some gray levels in the original histogram is greater than the threshold T, the value at that location is set to T; when the value of some gray levels in the original histogram is not greater than the threshold T, the value at that location remains unchanged, that is:

[0017]

[0018] Where T is the set threshold, k is the gray level of the image, 0≤k≤255, P r (k) is the original histogram, P T (k) is the platform histogram;

[0019] Step 203: Accumulate the calculation platform histogram to obtain the cumulative function

[0020]

[0021] Among them, F T (k) is the cumulative histogram of the image;

[0022] Step 204: Redistribute the grayscale of the image by accumulating the histogram to obtain a balanced image.

[0023]

[0024] Among them, R T (k) is the gray value of the pixel with gray level k after platform histogram equalization, 0≤R T (k)≤255, Indicates rounding;

[0025] Step 205: According to the formula y=cx γ Perform gamma transform on the equalized image, where c and γ are both positive numbers, x is the grayscale value of the input image, and y is the grayscale value of the output image.

[0026] In the above-mentioned image processing method for vehicle-mounted infrared pedestrian detection and tracking, the specific process of using HOG features and HOI features to extract pedestrian features from the enhanced vehicle-mounted infrared image in step 3 includes:

[0027] HOG feature extraction:

[0028] Step A1, Gamma normalization;

[0029] Step A2, gradient calculation;

[0030] Step A3, constructing a gradient direction histogram;

[0031] Step A4: normalization within the block;

[0032] Step A5, collecting HOG features;

[0033] HOI feature extraction:

[0034] Step B1, dividing the image into several blocks;

[0035] Step B2: normalization within the block;

[0036] Step B3: concatenate the HOI feature vectors in each block.

[0037] In the above-mentioned image processing method for vehicle-mounted infrared pedestrian detection and tracking, the specific process of optimizing the support vector machine parameters by the improved particle swarm algorithm in step 4 includes:

[0038] Step 401, initializing a particle swarm, forming a large number of initial particles and constituting an initial particle population, and assigning initial values ​​to penalty factors and kernel parameters in a support vector machine;

[0039] Step 402: Evaluate each particle in the particle population to obtain a fitness value of each particle;

[0040] Step 403: compare the fitness value of each particle with the fitness value of the particle's best historical position. When the current fitness value is better than the fitness value of the particle's best historical position, update the fitness value of the particle.

[0041] Step 404: compare the fitness value of each particle with the fitness value of the best historical position of the entire group. When the current fitness value is better than the fitness value of the best historical position of the entire group, update the fitness value of the particle.

[0042] Step 405, calculating the value of the inertia weight;

[0043] Step 406: Update the speed and position of each particle;

[0044] Step 407: When the maximum number of iterations is met or the optimal solution is found, execute step 408 to end the iteration; otherwise, execute step 402 to continue solving;

[0045] Step 408: Output the penalty factor, the global optimal value of the kernel parameter and the classification accuracy.

[0046] In the above-mentioned image processing method for vehicle-mounted infrared pedestrian detection and tracking, the fitness value in step 402, step 403 and step 404 is the classification accuracy of pedestrians in the support vector machine, and the formula is:

[0047]

[0048] In the formula, R is the classification accuracy, n totall is the total number of pedestrian samples, n correct are the pedestrian samples that are correctly classified.

[0049] In the above-mentioned image processing method for vehicle-mounted infrared pedestrian detection and tracking, the calculation formula of the inertia weight in step 405 is:

[0050]

[0051] Where w is the inertia weight, w maxis the maximum value of the inertia weight, η is the control factor, t is the number of iterations, t max is the maximum number of iterations, and σ is the fluctuation degree of the inertia weight.

[0052] In the above-mentioned image processing method for vehicle-mounted infrared pedestrian detection and tracking, the specific process of the scale-adaptive kernel correlation filter tracking algorithm in step 5 includes:

[0053] Step 501: input a vehicle-mounted infrared image sequence, detect pedestrians in the sequence, obtain an image block at a target position, and train a classifier to obtain initial position filter coefficients and scale filter parameters;

[0054] Step 502: Calculate the position-dependent filter response, and determine the target position of the current image according to the maximum value of the response;

[0055] Step 503: extract multi-scale samples with the current predicted target position as the center;

[0056] Step 504, calculating the scale kernel correlation filter response, the maximum value is the desired scale;

[0057] Step 505: Calculate the position of the tracking frame according to the predicted target position and the predicted scale;

[0058] Step 506: Update the filter model parameters and the target model and output the tracking result.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1. The method of the present invention has simple steps and is easy to implement.

[0061] 2. The present invention adopts median filtering to reduce infrared image noise, and adopts platform histogram and gamma transform to enhance infrared image contrast, thereby solving the problems of low signal-to-noise ratio and poor contrast of vehicle-mounted infrared images.

[0062] 3. The present invention uses directional gradient histogram and brightness histogram to extract infrared pedestrian features, adopts a support vector machine model optimized by particle swarm optimization, and optimizes the parameters of the particle swarm algorithm using a strategy of nonlinear decreasing inertia weight to address the problem that the particle swarm algorithm is prone to fall into local optimality, thereby increasing the global and local optimization capabilities of the particle swarm algorithm, thereby detecting vehicle-mounted infrared pedestrians, improving classification accuracy, and having good generalization ability, which can improve the accuracy of pedestrian detection.

[0063] 4. The present invention introduces a scale filter on the basis of the kernel correlation filter tracking algorithm to obtain a scale-adaptive kernel correlation filter tracking algorithm, thereby realizing scale-adaptive vehicle-mounted infrared pedestrian tracking, which can achieve better tracking results and meet real-time requirements.

[0064] 5. The present invention can be applied in automobile assisted driving systems to achieve accurate detection and stable tracking of vehicle-mounted infrared pedestrian targets, improve the active safety of vehicles, have significant effects, and be easy to promote.

[0065] In summary, the method of the present invention has simple steps and is easy to implement. It can be applied in automobile assisted driving systems to achieve accurate detection and stable tracking of vehicle-mounted infrared pedestrian targets, improve the active safety of vehicles, have significant effects, and is easy to promote.

[0066] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method of the present invention;

[0068] Figure 2 It is a flow chart of the particle swarm optimization support vector machine algorithm in the method of the present invention;

[0069] Figure 3 The present invention is a flowchart of the mid-scale adaptive kernel correlation filter tracking algorithm. DETAILED DESCRIPTION

[0070] like Figure 1 As shown, the image processing method for vehicle-mounted infrared pedestrian detection and tracking of the present invention comprises the following steps:

[0071] Step 1: Use median filtering to filter the vehicle-mounted infrared image;

[0072] In specific implementation, median filtering is a nonlinear filtering, which is effective in eliminating impulse noise. It can also prevent the edge of the image from being blurred while filtering out noise. Specifically, median filtering uses a sliding window and replaces the value of the window center point with the median of each point in the window.

[0073] Step 2: Use platform histogram equalization and gamma transform to enhance the filtered vehicle-mounted infrared image;

[0074] In specific implementation, image enhancement makes the contour, contrast and other characteristic information of the target in the image stand out, thereby strengthening useful information and weakening or eliminating useless information to a certain extent. The visual effect is better corrected, and finally an image that is more suitable for analysis and processing by humans or machines is obtained.

[0075] Step 3: Use HOG features and HOI features to extract pedestrian features from the enhanced vehicle-mounted infrared image to obtain edge information and brightness information of pedestrians in the vehicle-mounted infrared image;

[0076] Step 4: Use the improved particle swarm algorithm to optimize the support vector machine parameters for classification and obtain the specific position of the pedestrian in the vehicle-mounted infrared image;

[0077] Step 5: Use the scale-adaptive kernel correlation filter tracking algorithm to automatically track pedestrians in the vehicle-mounted infrared image.

[0078] In this embodiment, the specific process of platform histogram equalization and gamma transformation in step 2 includes:

[0079] Step 201: Determine the threshold value T by taking the average value

[0080]

[0081] Among them, f(x,y) is the filtered vehicle-mounted infrared image, M and N are the sizes of the image;

[0082] Step 202: When the value of some gray levels in the original histogram is greater than the threshold T, the value at that location is set to T; when the value of some gray levels in the original histogram is not greater than the threshold T, the value at that location remains unchanged, that is:

[0083]

[0084] Where T is the set threshold, k is the gray level of the image, 0≤k≤255, P r (k) is the original histogram, P T (k) is the platform histogram;

[0085] Step 203: Accumulate the calculation platform histogram to obtain the cumulative function

[0086]

[0087] Among them, F T (k) is the cumulative histogram of the image;

[0088] Step 204: Redistribute the grayscale of the image by accumulating the histogram to obtain a balanced image.

[0089]

[0090] Among them, R T (k) is the gray value of the pixel with gray level k after platform histogram equalization, 0≤R T (k)≤255, Indicates rounding;

[0091] Step 205: According to the formula y=cx γ Perform gamma transform on the equalized image, where c and γ are both positive numbers, x is the grayscale value of the input image, and y is the grayscale value of the output image.

[0092] In specific implementation, the gamma transform enlarges the mapping of the grayscale area. When γ=1, the gamma transform is transformed into a linear transform; when γ<1, the range of low grayscale levels of the image is increased, the range of high grayscale levels is reduced, and the image appears brighter; when γ>1, the range of high grayscale levels is increased, the range of low grayscale levels is reduced, and the image becomes darker.

[0093] In this embodiment, the specific process of using HOG features and HOI features to extract pedestrian features from the enhanced vehicle-mounted infrared image in step 3 includes:

[0094] HOG feature extraction:

[0095] Step A1, Gamma normalization;

[0096] In specific implementation, the image acquisition environment is different, the lighting environment is different, and the brightness of the obtained image is different. In order to make pedestrian detection stable and adapt to different environments, the image is corrected. In infrared images, Gamma normalization is used to adjust the image contrast, reduce noise interference, and weaken the influence of lighting and local image shadows on feature extraction. Camma correction formula:

[0097] I(x,y)=I(x,y) gamma

[0098] Where I(x,y) is the pixel value of the image and gamma is 1 / 2.

[0099] Step A2, gradient calculation;

[0100] In specific implementation, the gradient calculation method of the horizontal and vertical coordinate directions of each pixel point of the image is: using the one-dimensional gradient operator [-1 0 1] and [-1 0 1]T to perform convolution operations. The specific formula is:

[0101] G x (x,y)=I(x+1,y)-I(x-1,y)

[0102] G y (x,y)=I(x,y+1)-I(x,y-1)

[0103] In the formula, x represents the horizontal coordinate of the pixel point (x, y), y represents the vertical coordinate of the pixel point (x, y), G x (x,y),G y (x,y) represents the horizontal gradient and vertical gradient of the pixel point (x,y), respectively. I(x,y+1), I(x,y-1), I(x-1,y), and I(x+1,y) are the values ​​of the top, bottom, left, and right pixels, respectively.

[0104] The gradient magnitude G(x,y) and gradient direction θ(x,y) of the pixel are calculated based on the horizontal and vertical gradients. The specific formula is:

[0105]

[0106]

[0107] Where θ(x,y) ranges from [0°,180°].

[0108] Step A3, constructing a gradient direction histogram;

[0109] In the specific implementation, in order to make the image more robust when the appearance and posture of pedestrians in the image change slightly, the local gradient direction is used to describe it. The specific method is: divide the image into small connected areas, namely cells. A cell refers to a pixel area composed of several spatially adjacent pixels in the image. The gradient direction histogram of each pixel in the cell is collected. The spatially connected pixel area composed of each cell is called a block. A block contains a combination of feature vectors of the cells, and the HOG feature of the block is obtained.

[0110] Step A4: normalization within the block;

[0111] In the specific implementation, the local illumination changes and the difference in foreground-background contrast of different pictures make the gradient intensity vary greatly. The gradient intensity is normalized. The specific method is: the feature vector of each block is normalized using L2-norm normalization.

[0112] Step A5, collecting HOG features;

[0113] In the specific implementation, the feature vectors of all overlapping blocks in the detection window are finally collected and combined into the HOG feature for classification. The dimension of the HOG feature descriptor can be obtained by the selection of cells and blocks. The number of cells in each block is represented as β, the number of blocks in each detection window is represented as γ, and the number of histogram channels (bins) in each cell is represented as η. The dimension of the final feature vector can be expressed as β×γ×η.

[0114] HOI feature extraction:

[0115] Step B1, dividing the image into several blocks;

[0116] In specific implementation, the detection window is divided into small areas, namely cells. For each cell, the grayscale value of each pixel in the cell is calculated to form a local one-dimensional brightness histogram. In each brightness histogram, the grayscale value is mapped to different intervals, and the size of the grayscale value is mapped to the interval as a weight value.

[0117] Step B2: normalization within the block;

[0118] In the specific implementation, cells are overlapped between blocks, so that each independent cell information can be shared in several different blocks, and then the histogram in the block area is normalized to improve the robustness of HOI for pedestrian description.

[0119] Step B3: concatenate the HOI feature vectors in each block.

[0120] In this embodiment, Figure 2 As shown, the specific process of optimizing the support vector machine parameters by the improved particle swarm algorithm described in step 4 includes:

[0121] Step 401, initializing a particle swarm, forming a large number of initial particles and constituting an initial particle population, and assigning initial values ​​to penalty factors and kernel parameters in a support vector machine;

[0122] Step 402: Evaluate each particle in the particle population to obtain a fitness value of each particle;

[0123] Step 403: compare the fitness value of each particle with the fitness value of the particle's best historical position. When the current fitness value is better than the fitness value of the particle's best historical position, update the fitness value of the particle.

[0124] Step 404: compare the fitness value of each particle with the fitness value of the best historical position of the entire group. When the current fitness value is better than the fitness value of the best historical position of the entire group, update the fitness value of the particle.

[0125] Step 405, calculating the value of the inertia weight;

[0126] Step 406: Update the speed and position of each particle;

[0127] Step 407: When the maximum number of iterations is met or the optimal solution is found, execute step 408 to end the iteration; otherwise, execute step 402 to continue solving;

[0128] Step 408: Output the penalty factor, the global optimal value of the kernel parameter and the classification accuracy.

[0129] In the specific implementation, after obtaining the binary classifier through the particle swarm optimization support vector machine algorithm, in order to determine the specific location of the pedestrian in the input infrared image, it is necessary to use a sliding window to scan the image. The sliding window traverses the detection image from left to right and from top to bottom in turn, and uses the trained support vector machine to perform classification and recognition in each corresponding window to determine whether the window contains pedestrians. Specifically, the minimum size of the detection window is set, the entire image is scanned, and the detection window is scaled based on this, the scaling ratio is set to 1.2, the image is scanned repeatedly, and pedestrian features are extracted and classified for each corresponding window to determine whether it contains pedestrians. Finally, the detection results of each scale are fused.

[0130] In this embodiment, the fitness values ​​in step 402, step 403 and step 404 are the classification accuracy of pedestrians in the support vector machine, and the formula is:

[0131]

[0132] In the formula, R is the classification accuracy, n totall is the total number of pedestrian samples, n correct are the pedestrian samples that are correctly classified.

[0133] In this embodiment, the calculation formula of the inertia weight in step 405 is:

[0134]

[0135] Where w is the inertia weight, w max is the maximum value of the inertia weight, η is the control factor, t is the number of iterations, t max is the maximum number of iterations, and σ is the fluctuation degree of the inertia weight.

[0136] In the specific implementation, the value of η is 0.81 and the value of σ is 0.065. An exponential function is used to control the change of w. As the number of iterations increases, the nonlinearity of the exponential function decreases. In the later stage of iteration, the sine function increases the global search ability of the algorithm and reduces the possibility of the algorithm falling into the local optimum.

[0137] In this embodiment, Figure 3 As shown, the specific process of the scale-adaptive kernel correlation filter tracking algorithm in step 5 includes:

[0138] Step 501: input a vehicle-mounted infrared image sequence, detect pedestrians in the sequence, obtain an image block at a target position, and train a classifier to obtain initial position filter coefficients and scale filter parameters;

[0139] Step 502: Calculate the position-dependent filter response, and determine the target position of the current image according to the maximum value of the response;

[0140] Step 503: extract multi-scale samples with the current predicted target position as the center;

[0141] Step 504, calculating the scale kernel correlation filter response, the maximum value is the desired scale;

[0142] Step 505: Calculate the position of the tracking frame according to the predicted target position and the predicted scale;

[0143] Step 506: Update the filter model parameters and the target model and output the tracking result.

[0144] In specific implementation, when the vehicle continues to travel, the size of the pedestrian will change with the distance from the vehicle, the size of the pedestrian target will change, and there will be a deviation between the size of the tracking window and the size of the actual target. In order to make the tracking suitable for an environment where the target scale changes, the present invention adopts a scale-adaptive kernel correlation filter tracking algorithm based on the kernel correlation filter tracking algorithm, introduces target scale estimation, establishes training samples of different scales around the target position, trains a scale filter, obtains the scale with the largest response, and realizes the scale update.

[0145] In order to verify the effect of the method of the present invention, distance precision (DP), overlap success (OS) and frame rate (FPS) are used for quantitative comparative analysis. Distance precision refers to the percentage of frames in which the center position error of all images is less than the given center position error threshold in the total number of frames. Distance precision is described by center location error (CLE). CLE represents the Euclidean distance between the center position of the tracked target and the center position of the actual target. The smaller the center position error, the better the tracking effect. The deviation threshold is selected as 20 pixels during verification. The success rate is expressed by the overlap rate, which refers to the ratio of the overlapping area of ​​the tracked target rectangular frame and the actual target rectangular frame to the total area. When the overlap rate of each frame is higher than the given threshold, it means that the frame is successfully tracked. The percentage of the number of frames that are successfully tracked in the total number of frames is the success rate. For vehicle-mounted infrared pedestrian tracking, real-time performance is very important. The frame rate is an indicator that characterizes real-time performance, that is, the number of frames that the tracking algorithm can process per unit time.

[0146] Specifically, continuous images are selected from the LSI far-infrared pedestrian dataset for testing. The kernel correlation filter tracking algorithm and the scale-adaptive kernel correlation filter tracking algorithm are used to detect and track pedestrians that are gradually decreasing and to detect and track pedestrians that are gradually increasing. The calculated distance accuracy comparison is shown in Table 1, the success rate comparison is shown in Table 2, and the frame rate comparison is shown in Table 3.

[0147] Table 1 Distance accuracy comparison

[0148]

[0149] Table 2 Comparison of success rates

[0150]

[0151] Table 3 Frame rate comparison (f / s)

[0152]

[0153] As can be seen from Tables 1 and 2, the distance accuracy and success rate of the scale-adaptive kernel correlation filter tracking algorithm are 0.5% and 80% higher than those of the kernel correlation filter tracking algorithm in an environment where pedestrians are gradually getting smaller, and 25.3% and 66.7% higher in an environment where pedestrians are gradually getting larger, respectively, achieving better tracking results. It can be seen from Table 3 that the speed of the present invention is greatly reduced, and its frame rate is reduced to 30.69s and 125.26s, respectively, but it can still meet the real-time requirements of vehicle-mounted infrared pedestrian tracking.

[0154] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An image processing method for vehicle-mounted infrared pedestrian detection and tracking, characterized in that: The following steps are involved: Step 1: Use median filtering to filter the vehicle-mounted infrared image; Step 2: Use platform histogram equalization and gamma transform to enhance the filtered vehicle-mounted infrared image; The specific process of the platform histogram equalization and gamma transformation includes: Step 201: Determine the threshold value T by taking the average value Among them, f(x,y) is the filtered vehicle-mounted infrared image, M and N are the sizes of the image; Step 202: When the value of some gray levels in the original histogram is greater than the threshold T, the value at that location is set to T; when the value of some gray levels in the original histogram is not greater than the threshold T, the value at that location remains unchanged, that is: Where T is the set threshold, k is the gray level of the image, 0≤k≤255, P r (k) is the original histogram, P T (k) is the platform histogram; Step 203: Accumulate the calculation platform histogram to obtain the cumulative function Among them, F T (k) is the cumulative histogram of the image; Step 204: Redistribute the grayscale of the image by accumulating the histogram to obtain a balanced image. Among them, R T (k) is the gray value of the pixel with gray level k after platform histogram equalization, 0≤R T (k)≤255, Indicates rounding; Step 205: According to the formula y=cx γ Perform gamma transformation on the equalized image, where c and γ are both positive numbers, x is the grayscale value of the input image, and y is the grayscale value of the output image; Step 3: Use HOG features and HOI features to extract pedestrian features from the enhanced vehicle-mounted infrared image to obtain edge information and brightness information of pedestrians in the vehicle-mounted infrared image; Step 4: Use the improved particle swarm algorithm to optimize the support vector machine parameters for classification and obtain the specific position of the pedestrian in the vehicle-mounted infrared image; Step 5: Use the scale-adaptive kernel correlation filter tracking algorithm to automatically track pedestrians in the vehicle-mounted infrared image.

2. The image processing method for vehicle-mounted infrared pedestrian detection and tracking according to claim 1, characterized in that: The specific process of using HOG features and HOI features to extract pedestrian features from the enhanced vehicle-mounted infrared image in step 3 includes: HOG feature extraction: step A1, Gamma normalization; Step A2, gradient calculation; Step A3, constructing a gradient direction histogram; Step A4: normalization within the block; Step A5, collecting HOG features; HOI feature extraction: Step B1, dividing the image into several blocks; Step B2: normalization within the block; Step B3: concatenate the HOI feature vectors in each block.

3. The image processing method for vehicle-mounted infrared pedestrian detection and tracking according to claim 1, characterized in that: The specific process of optimizing the support vector machine parameters by the improved particle swarm algorithm described in step 4 includes: Step 401, initializing a particle swarm, forming a large number of initial particles and constituting an initial particle population, and assigning initial values ​​to penalty factors and kernel parameters in a support vector machine; Step 402: Evaluate each particle in the particle population to obtain a fitness value of each particle; Step 403: compare the fitness value of each particle with the fitness value of the particle's best historical position. When the current fitness value is better than the fitness value of the particle's best historical position, update the fitness value of the particle. Step 404: compare the fitness value of each particle with the fitness value of the best historical position of the entire group. When the current fitness value is better than the fitness value of the best historical position of the entire group, update the fitness value of the particle. Step 405, calculating the value of the inertia weight; Step 406: Update the speed and position of each particle; Step 407: When the maximum number of iterations is met or the optimal solution is found, execute step 408 to end the iteration; otherwise, execute step 402 to continue solving the problem; Step 408: Output the penalty factor, the global optimal value of the kernel parameter and the classification accuracy.

4. The image processing method for vehicle-mounted infrared pedestrian detection and tracking according to claim 3, characterized in that: The fitness values ​​in step 402, step 403 and step 404 are the classification accuracy of pedestrians in the support vector machine, and the formula is: In the formula, R is the classification accuracy, n totall is the total number of pedestrian samples, n correct are the pedestrian samples that are correctly classified.

5. The image processing method for vehicle-mounted infrared pedestrian detection and tracking according to claim 3, characterized in that: The calculation formula of the inertia weight in step 405 is: Where w is the inertia weight, w max is the maximum value of the inertia weight, η is the control factor, t is the number of iterations, t max is the maximum number of iterations, and σ is the fluctuation degree of the inertia weight.

6. The image processing method for vehicle-mounted infrared pedestrian detection and tracking according to claim 1, characterized in that: The specific process of the scale-adaptive kernel correlation filter tracking algorithm described in step 5 includes: Step 501: input a vehicle-mounted infrared image sequence, detect pedestrians in the sequence, obtain an image block at a target position, and train a classifier to obtain initial position filter coefficients and scale filter parameters; Step 502: Calculate the position-dependent filter response, and determine the target position of the current image according to the maximum value of the response; Step 503: extract multi-scale samples with the current predicted target position as the center; Step 504, calculating the scale kernel correlation filter response, the maximum value is the desired scale; Step 505: Calculate the position of the tracking frame according to the predicted target position and the predicted scale; Step 506: Update the filter model parameters and the target model and output the tracking result.