Improved carnivorous plant algorithm optimized vehicle long-term target tracking method

Abstract

The application discloses a kind of improved long-term target tracking methods of vehicle of carnivorous plant algorithm optimization, extract tracking area direction gradient histogram as feature template, based on a kind of carnivorous plant algorithm tracking framework of Bart distance similarity function design;Second, a short-term memory module is designed to predict the initialization individual of carnivorous plant algorithm in tracking process, simultaneously, a random following strategy and population division mechanism are designed in the iteration process, and are introduced into the tracking framework as search strategy;On the basis of direction gradient histogram, the two layers of features in Conv2-1 and Conv4-3 in ResNet-50 are fused, an adaptive learning rate dynamic updating template is designed, a two-dimensional scale perception factor is added in the population dimension to adapt to the scale change of target window.The experimental results show that the application can adapt to the change of environment in complex scene, and keep stable tracking for target vehicle for a long time.Compared with the prior art, the tracking accuracy and robustness are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of target tracking technology, and more specifically to a long-term vehicle target tracking method optimized by an improved carnivorous plant algorithm. Background Technology

[0002] Unmanned aerial vehicle (UAV) vehicle tracking is an important area of ​​computer vision research. It not only tracks vehicle trajectories from the air but also provides a reliable source of scientific data for intelligent transportation, holding significant practical value in traffic monitoring and detection. The tracking process involves predicting the target's position and scale in each subsequent frame based on the target's position in the first frame. However, due to external factors such as lighting changes, interference from similar objects, and occlusion from the UAV's perspective, as well as the impact of target attitude changes such as rapid movement and rotation, tracking performance is often poor. Therefore, designing high-precision and real-time target tracking methods is crucial to address these issues. Current mainstream tracking models can be categorized into deep learning and correlation filtering. Correlation filtering tracking algorithms are robust and fast. Since the MOSSE correlation filtering tracking algorithm was proposed in 2010, various correlation tracking algorithms such as CN, KCF, SRDCF, FDSST, and ECO have emerged. However, these algorithms have limited tracking performance in situations involving rapid movement, occlusion, and deformation, and further improvements are needed. Since 2014, deep learning tracking algorithms have gradually become the mainstream of academic research. DLT was the first to introduce deep learning into the field of object tracking, and other well-known tracking algorithms include MDNet and SiamFC. However, deep learning tracking algorithms are highly dependent on datasets and high-performance computing platforms, excessively consuming computational resources. In recent years, with the rapid development of Swarm Intelligence (SI) algorithms and their demonstrated good optimization capabilities, target tracking algorithms optimized based on SI have become a new research hotspot. Ong KM et al. used an improved flower pollination algorithm to locate moving objects using a search window, but this algorithm relies solely on extracting the HSV histogram of the object's appearance, resulting in weak target representation. Nenavath et al. proposed a tracking algorithm based on a hybrid of teaching optimization and sine / cosine optimization, which showed good optimization ability and tracking performance, but poor scale adaptability. Zhang et al. proposed a hybrid algorithm based on extended ant colony optimization and sine / cosine optimization, which is highly competitive in tracking sudden movements, but this method relies on a large number of iterations to ensure tracking accuracy, leading to significant time consumption. Ong P. et al. used a pollination algorithm to track athletes, showing good performance in terms of tracking accuracy and processing time, but poor tracking results under occlusion conditions. These findings demonstrate the feasibility of swarm intelligence algorithms in optimizing target tracking. The Carnivorous Plant Algorithm (CPA), proposed by Ong KM in 2020, is a novel swarm intelligence algorithm characterized by fewer control parameters, strong adaptability, easy model modification, strong robustness, and ease of implementation. CPA can be applied to fields such as optimal controller design and image processing, but its optimization capabilities still have room for improvement. Summary of the Invention

[0003] The purpose of this invention is to provide an improved long-term vehicle target tracking method optimized by the carnivorous plant algorithm, in order to solve the tracking failure problem caused by complex situations such as cluttered backgrounds, rapid movement, and occlusion during the UAV vehicle tracking process. This invention designs a CPA method (CPA with RandomFollow Strategy-population Division Mechanism, RDCPA) based on a random follow strategy and population partitioning mechanism to optimize the UAV vehicle target tracking process, which can effectively improve tracking accuracy.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] An improved long-term vehicle target tracking method optimized by the carnivorous plant algorithm includes the following steps:

[0006] Step 1: Locate and track the target from the t-th frame of the vehicle video sequence, obtain the target position information in the t-th frame, extract features from the tracked target, and obtain the target template γ(t) in the t-th frame;

[0007] Step 2: Based on the target template, use the improved carnivorous plant algorithm to search for several target candidate boxes;

[0008] Step 3: Generate several candidate templates from the (t+1)th frame of the vehicle video sequence based on several target candidate boxes;

[0009] Step 4: Perform similarity matching between the candidate template extracted in frame t+1 and the target template extracted in frame t. Select the target candidate box corresponding to the candidate template with the best similarity as the tracking target in frame t+1. Extract features from the tracking target to obtain the target template γ in frame t+1. best (t+1);

[0010] Step 5: Fuse the target template of frame t+1 and the target template of frame t through a template learning mechanism to obtain a new target template γ(t+1). Repeat steps 2-5 until the last frame of the video sequence.

[0011] Furthermore, the feature extraction of the tracked target specifically includes:

[0012] The FHOG features and deep features of the tracked target are weighted and fused. The deep features are extracted using the ResNet-50 network model, which achieves 75% success rate in ImageNet training. The deep features include Conv2-1 and Conv4-3 features from the ResNet-50 network model. Therefore, the target template γ(t) is represented as follows:

[0013] γ(t)=λ1γ Conv2-1 (t)+λ2γConv4-3 (t)+λ3γ FHOG (t)

[0014] Where, γ Conv2-1 γ Conv4-3 and γ FHOG λ1, λ2, and λ3 represent the Conv2-1, Conv4-3, and FHOG features of the tracked target, respectively, and λ1, λ2, and λ3 are feature fusion coefficients.

[0015] Furthermore, λ1 = 0.8, λ2 = 0.5, and λ3 = 0.5.

[0016] Furthermore, the step of searching for several target candidate boxes based on the target template using the improved carnivorous plant algorithm specifically includes:

[0017] Step 2.1: Based on the target location information in frame t, obtain several target location information in frame t+1 through the short-term memory module;

[0018] Step 2.2: Based on the location information of the short-term memory module, use the improved carnivorous plant algorithm to search for more target location information and obtain the location of the target candidate box;

[0019] Step 2.3: Based on the improved carnivorous plant algorithm, obtain more target location information, use the scale perception factor at each target location to obtain the target scale information, and determine the size of the target candidate box.

[0020] Furthermore, the short-term memory module adopts a Gaussian distribution method, uses the position of the target in three frames of the video sequence to predict the motion trajectory, allows the population initialization to be as close as possible to the predicted target area, makes full use of the individual's search ability, shortens the global optimization time, and locates the local target faster.

[0021] Individual X i It follows a multidimensional Gaussian distribution, as shown in the following formula:

[0022]

[0023] Where p(X) i (t+1) represents individual X i The probability of occurrence, σ t+1 and μ t+1 Here, we have the covariance and mean matrix of the initial population at time t+1, respectively. The mean matrix is ​​related to the predicted target position and is calculated using the following formula:

[0024] μ t+1 =[X best (t)-X best (t-1),X best (t-1)-Xbest (t-2)]

[0025] In the formula, X best The optimal solutions obtained by the CPA algorithm for frames t, t-1, and t-2;

[0026] In the short-term memory module, the target position obtained from the previous three frames is used to determine the motion velocity in the current frame, which is then used as the mean μ of the Gaussian model. t+1 Calculate the initial position of the population in the current frame, and use it as the target position information for the (t+1)th frame.

[0027] Furthermore, the reproduction process of the improved carnivorous plant algorithm is as follows:

[0028] C i (k)=C i (k-1)+rr×exp(f(C i (k-1))-f(C v (k-1)))×M i

[0029]

[0030] Among them, C i Let P represent the plant in the carnivorous plant algorithm. i For prey, i≠v, C v For randomly selected carnivorous plants, r r Let f be the reproduction rate, set to 1.8, f represent the similarity of individuals, and k represent the current iteration number of the CPA algorithm;

[0031] If the population size is n and the number of plants is n c The remaining individuals are all prey, and their number is n. p By setting a population division mechanism, the number of prey iterations is adaptively reduced, and the number of plants is adaptively increased with the number of iterations. Therefore, a proportional factor κ is designed to adjust the population ratio, with the number of plants being n×κ and the number of prey being n×(1-κ).

[0032]

[0033] Among them, κ first and κ end Let κ1 be the initial and final values ​​of the proportional gain, set to 0.2 and 0.5 respectively, κ1 be the damping coefficient (greater than 0), and k be the current iteration number. max This represents the maximum number of iterations.

[0034] Furthermore, the scale sensing factor s i =[s xi s yi ], s xi and syi Let X be the scaling factor for the length and width of the target candidate box, respectively. Then the new CPA individual X... i =[x i y i s xi s yi The FHOG features of each scale window are extracted, and the response values ​​corresponding to different scale factors are calculated. The optimal scale response is selected to complete the scale update. It is assumed that the target window size for the current frame is M. O ×N O The current frame image size is M P ×N P Therefore, the boundary of the proportionality perception factor is:

[0035] LB s =τ s ×exp(log(max(5 / M O ,5 / N O )) / log(τ s ))

[0036] UB s =τ s ×exp(log(min(M P / M O M P / N O )) / log(τ s ))

[0037] Among them, LB s UB is the lower bound of the scale perception factor. s τ is the upper bound of the scale perception factor. s The scale step size is set to 1.02;

[0038] To accelerate tracking, the scale is updated every 10 frames along with the position update, and each scale update extracts n×k values. max Different scales of samples, where n refers to the population size, k max This indicates the maximum number of iterations for the CPA algorithm.

[0039] Furthermore, the similarity matching in step 4 specifically involves:

[0040] The Bhattacharyya coefficients are used as the solution condition for template matching. The formula for the Bhattacharyya coefficients is as follows:

[0041]

[0042] In the formula, a and b are feature matrices, z is the dimension of the feature matrix, and ρ is the obtained Bhattacharyya coefficient, which ranges from [0,1]. The closer it is to 1, the higher the similarity between features.

[0043] Furthermore, in step 5, fusion is achieved through a template learning mechanism, specifically as follows:

[0044] Choosing an appropriate learning rate ensures the stability of long-term tracking. The optimal target template matched in frame t+1 is γ. best (t+1), whose corresponding individual fitness value is F best (t+1), then the adaptive learning rate for the (t+1)th frame is:

[0045]

[0046] Where, Φ t+1 Let Φ be the learning rate for frame t+1. max The peak value of the learning rate is δ, and ξ are adjustment parameters. best (t+1) represents the optimal fitness value of the population after the tracking algorithm finishes tracking in frame t+1;

[0047] The target template of frame t+1 and the target template of frame t are fused together using a template learning mechanism to obtain a new target template:

[0048] γ(t+1)=(1-φ t+1 )·γ(t)+φ t+1 ·γ best (t+1)

[0049] Where γ(t) is the target template of the t-th frame, γ best (t+1) is the target template for the (t+1)th frame.

[0050] Furthermore, Φ max Set it to 0.04, and set δ and ξ to 10 and 5 respectively.

[0051] Compared with the prior art, the present invention has the following beneficial technical effects:

[0052] This invention investigates the fusion of swarm intelligence algorithms and target tracking. Using FHOG features as a template model and Bach distance as the similarity function, a carnivorous plant algorithm tracking framework is designed. The iterative method of CPA is improved by fusing deep features, employing an adaptive learning rate to dynamically update template features, and incorporating a scale factor to adapt to changes in window size, ensuring both accuracy and real-time performance in complex environments. Experimental results using vehicle video sequences selected from the UAV123 dataset demonstrate that the proposed method exhibits strong feature representation capabilities under changing vehicle scene conditions, achieving good tracking success rate and accuracy. Verification on aerial vehicle videos shows that the invention demonstrates good adaptability under complex conditions such as vehicle angle changes, low resolution, and strong light occlusion. Attached Figure Description

[0053] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0054] Figure 1 This is a flowchart illustrating the present invention;

[0055] Figure 2 The following are the tracking effects of the present invention on aerial vehicle videos, where (a) is frame 74, (b) is frame 472, (c) is frame 509, (d) is frame 520, (e) is frame 1110, and (f) is frame 1150. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] See Figure 1 This is a flowchart of the invention. The invention uses Fast Oriented Gradient Histogram (HGP) features as a template model and designs a carnivorous plant tracking algorithm framework using Bhattacharyya distance as the similarity function. It also improves the CPA iteration method, integrates deep features, employs an adaptive learning rate to dynamically update the template features, and incorporates a scale factor to adapt to changes in window size, ensuring both accuracy and real-time performance in complex environments. Figure 2 As can be seen, during the tracking process, the exact position of the target vehicle in the image sequence is calculated using this invention. When passing through traffic lights, strong light occlusion causes a sudden and significant reduction in target information; this invention's algorithm overcomes this strong light interference and continues to track the vehicle. In frame 1110, encountering prolonged tree occlusion, the short-term memory module accurately predicts the tracking position and continues tracking the vehicle after the occlusion ends. In contrast, the compared FDSST algorithm loses track of the target. Qualitative analysis further verifies that this invention's algorithm has good tracking accuracy and robustness when dealing with various interference factors such as similar targets, shrub occlusion, changes in angle and scale, and image blur. Especially when the vehicle passes through traffic lights, the drone's tracking and shooting angle is affected by strong light and occlusion, resulting in a significant reduction in information; this invention demonstrates superior performance.

[0059] The specific steps are as follows:

[0060] Step 1: Locate and track the target from the t-th frame of the vehicle video sequence, obtain the target position information in the t-th frame, extract features from the tracked target, and obtain the target template γ(t) in the t-th frame;

[0061] Feature extraction transforms target information into digital information, using matrices or vectors to perform spatial mapping, extracting target features from the first frame, and searching for the best-matching target in subsequent frames. FHOG (Fast Histogram of Oriented Gradients) combines unsigned and signed directional gradient histograms of local regions to represent features. First, orientation-sensitive and orientation-insensitive regions of the target candidate area are mapped, normalized, and truncated. Then, 18 dimensions of orientation-sensitive features and 9 dimensions of orientation-insensitive features are statistically obtained, along with 4 accumulated features, for a total of 31 dimensions. FHOG features can describe the edge information of the tracked target in detail, using local shape features to describe information, and are adaptable to changes in illumination intensity and target shape. It exhibits good robustness in representing features of pedestrians, vehicles, and other objects in UAV scenarios.

[0062] Deep features possess superior local texture representation capabilities, contain more semantic information, and can better identify target deformations. Therefore, a weighted fusion of FHOG features and deep features is proposed. A convolutional neural network is intended to extract deep features of the target region. Many convolutional neural network models have emerged in recent years, among which the ResNet network learns the residual between input and output through multiple parametric layers to ensure model accuracy. Therefore, the ResNet network model is adopted for feature extraction of the tracked target.

[0063] A ResNet-50 network model, which achieved a 75% success rate in ImageNet training, was selected to extract depth features of the tracked target. Extensive experimental validation showed that the Conv4-3 layer contains more semantic information, reduces the impact of illumination changes, and effectively identifies target deformation. The Conv2-1 layer contains more detailed target information and is more effective in locating boundaries, enabling precise target localization. Therefore, this paper extracts features from both the Conv2-1 and Conv4-3 layers of the ResNet-50 network, along with the FHOG features of the target region. Thus, the target template γ(t) is represented as follows:

[0064] γ(t)=λ1γ Conv2-1 (t)+λ2γ Conv4-3 (t)+λ3γ FHOG (t)

[0065] Where, γ Conv2-1 γ Conv4-3 and γ FHOGλ1, λ2, and λ3 represent the Conv2-1, Conv4-3, and FHOG features of the tracked target, respectively, while λ1, λ2, and λ3 are feature fusion coefficients. During the experiments, it was found that different weights yielded different results. Through extensive experiments, the empirical values ​​for the fusion weights in the improved algorithm were determined to be: λ1 = 0.8, λ2 = 0.5, and λ3 = 0.5.

[0066] Step 2: Based on the target template, search for several target candidate boxes using the improved carnivorous plant algorithm; specifically, including:

[0067] Step 2.1: Based on the target location information in frame t, obtain several target location information in frame t+1 through the short-term memory module;

[0068] Step 2.2: Based on the location information of the short-term memory module, use the improved carnivorous plant algorithm to search for more target location information and obtain the location of the target candidate box;

[0069] Step 2.3: Based on the improved carnivorous plant algorithm, obtain more target location information, use the scale perception factor at each target location to obtain the target scale information, and determine the size of the target candidate box.

[0070] The short-term memory module includes:

[0071] Initialization affects the optimization results and efficiency of CPA (Collision Perception) tracking. If individuals are distributed across all searchable areas, real-time performance will suffer. Therefore, this paper designs a short-term memory module to improve the CPA initialization method. The module uses a Gaussian distribution and predicts the motion trajectory using the target's position in two frames of a video sequence. This allows the initial population to be as close as possible to the predicted target area, fully utilizing the individual's search capabilities, shortening the global optimization time, and locating the local target more quickly. Individual X i It follows a multidimensional Gaussian distribution, as shown in the following formula:

[0072]

[0073] Where p(X) i (t+1) represents individual X i The probability of occurrence, σ t+1 and μ t+1 Here, we have the covariance and mean matrix of the initial population at time t+1, respectively. The mean matrix is ​​related to the predicted target position and is calculated using the following formula:

[0074] μ t+1 =[X best (t)-X best (t-1),X best (t-1)-X best (t-2)]

[0075] In the formula, X best The optimal solutions obtained by the CPA algorithm for frames t, t-1, and t-2;

[0076] In the short-term memory module, the target position movement distance obtained from the previous two frames is used to determine the movement velocity of the model in the current frame. This velocity is then used as the mean of the Gaussian model to calculate the initial position of the population in the current frame. The short-term memory module helps to concentrate the population at the target location as much as possible, and joint random initialization ensures population diversity while reducing the computational cost of the CPA algorithm. Simultaneously, the short-term memory module stores the velocity information from the previous two frames, providing position prediction capabilities and mitigating risks and challenges such as occlusion, thus ensuring long-term tracking capability during the tracking process.

[0077] Basic carnivorous plant algorithms include:

[0078] Randomly initialize n individuals consisting of carnivorous plants and prey, with the number of carnivorous plants and prey denoted as n. C and n P The positions of each body are represented by a matrix as follows:

[0079]

[0080] x ij =LB + (UB - LB) × rand

[0081] In the formula: x ij Let be the position of the i-th plant or prey in j-dimensional space, n be the total population, d be the dimension of the problem to be solved, rand be a random number, and UB and LB be the upper and lower bounds of the problem to be solved. The fitness value of each carnivorous plant or prey is as follows:

[0082]

[0083] Each individual is sorted in ascending order of fitness value. The top n individuals of the sorted population... C One individual is a carnivorous plant, and the remaining individuals are prey. The grouping process in CPA is as follows: Figure 1 As shown. The prey with the best fitness value is assigned to the top-ranked carnivorous plant, and second- and third-ranked prey are assigned to the second- and third-ranked carnivorous plants, respectively. This process is repeated until the nth time. C The nth prey was assigned to the nth prey. C Among the carnivorous plants, the first column completes the grouping. Then, starting from the (n)th column... C +1) prey begins to repeat the process of the first column and completes the grouping of the second column, and finally repeats the grouping of each column until all prey are assigned.

[0084] After grouping, each group contains only one carnivorous plant, and the number of prey must be greater than two. During the growth process, an attraction rate index is defined to determine the growth of both the carnivorous plant and its prey, set to 0.8. If the attraction rate is higher than a random number, the prey is captured and digested by the carnivorous plant and then begins to grow. The growth model of the new carnivorous plant in the k-th iteration is as follows:

[0085] C i (k)=g C ×C i (k-1)+(1-g C )×P v (k-1)

[0086] g C =g r ×rand i

[0087] Among them, C i For a carnivorous plant of level i, P v g is the randomly selected prey. r The growth rate is set to 2 here, the same as in the original CPA paper. The growth processes of both carnivorous plants and prey are repetitive until a predefined number of updates, g, is reached. k The value is set to 2 in the text. On the other hand, if the attraction rate is lower than the generated random value, the prey will try to escape the trap, which is mathematically represented as:

[0088] P i (k)=g P ×P u (k-1)+(1-g P )×P v (k-1)

[0089]

[0090] Where u≠v, P u and P v These are two other prey randomly selected from the population.

[0091] Reproduction exists only in first-order carnivorous plants, which represents the optimal solution within the population. The reproductive process of carnivorous plants can be represented as follows:

[0092] C i (k)=C1(k-1)+r C ×rand i ×M i

[0093]

[0094] Where i≠v≠1, C1 is the optimal solution, C v For randomly selected carnivorous plants, rC For the reproduction rate, this paper uses 1.8, the same as the original CPA paper.

[0095] New carnivorous plants and prey combined with the previous population, resulting in [n+n] C (g k )+n C A new population. Specifically, n individuals, n... C (g k ) new individuals and n C The individuals are defined as those in the original population, those during the growth process, and those during the reproduction process. The top n individuals from this population are selected as new candidate solutions based on their fitness.

[0096] The improved carnivorous plant algorithm includes:

[0097] In the standard CPA algorithm's breeding process, population updates only occur in the first level of carnivorous plants. When plants move towards the optimal position, the population is prone to rapid aggregation within a short period. While this achieves fast convergence, it significantly increases the probability of the algorithm getting trapped in local optima. Therefore, the new breeding process is as follows:

[0098] C i (k)=C i (k-1)+r r ×exp(f(C i (k-1))-f(C v (k-1)))×M i

[0099] The random follow strategy allows plants to move closer to the next level of plants with a certain probability, which ensures the convergence speed without reducing population diversity and can well balance global search and local development.

[0100] In CPA, if the population size is n and the number of plants is n c The remaining individuals are all prey, and their number is n. p In the early stages of algorithm iteration, prey individuals facilitate global exploration. If the number of prey is small and the number of plants is large, global exploration will be insufficient, local development will be overly refined, and the algorithm will easily fall into local optima. Conversely, in the later stages of iteration, if there are few plants and a large number of prey, local development will be lacking, resulting in low optimization accuracy. To fully utilize the capabilities of each plant and prey individual, a population partitioning mechanism is designed, where the prey size adaptively decreases with the number of iterations, and the number of plants adaptively increases accordingly. Therefore, a scaling factor κ is designed to adjust the population proportion, with the number of plants being n×κ and the number of prey being n×(1-κ).

[0101]

[0102] Among them, κ first and κ end Let κ1 be the initial and final values ​​of the proportional gain, set to 0.2 and 0.5 respectively, κ1 be the damping coefficient (greater than 0), and k be the current iteration number. max This represents the maximum number of iterations.

[0103] Proportion perception factors include:

[0104] To ensure the long-term tracking capability of the algorithm, let the proportional sensing factor s be... i =[s xi s yi ], where are the scaling factors for the length and width of the target box, respectively. Then, the new CPA individual X i =[x i y i s xi s yi The FHOG features of each scale window are extracted, and the response values ​​corresponding to different scale factors are calculated. The optimal scale response is then selected to complete the scale update. Assume the target window size for the current frame is M. O ×N O The current frame image size is M P ×N P Therefore, the boundary of the proportionality perception factor is:

[0105] LB s =τ s ×exp(log(max(5 / M O ,5 / N O )) / log(τ s ))

[0106] UB s =τ s ×exp(log(min(M P / M O M P / N O )) / log(τ s ))

[0107] Where, τ s The scale step size is set to 1.02.

[0108] To accelerate tracking, the scale is updated every 10 frames along with the position update, and each scale update extracts n×k values. max Different scale samples were used to obtain the scale corresponding to the maximum response value.

[0109] Step 3: Generate several candidate templates from the (t+1)th frame of the vehicle video sequence based on several target candidate boxes;

[0110] Step 4: Perform similarity matching between the candidate template extracted in frame t+1 and the target template extracted in frame t. Select the target candidate box corresponding to the candidate template with the best similarity as the tracking target in frame t+1. Extract features from the tracking target to obtain the target template γ in frame t+1. best (t+1);

[0111] Similarity determines the degree of similarity between the current feature image and the template feature image. Considering that target tracking is primarily based on constrained optimization problems, and that Bhattacharyya distance (the Bhattacharyya coefficient) is an approximate estimate of the overlap between two sample vectors, which can be used to measure the relative closeness of two vectors, it is used as a condition for template matching. The formula for the Bhattacharyya coefficient is as follows:

[0112]

[0113] In the formula, a and b are feature matrices, z is the dimension of the feature matrix, and ρ is the obtained Bhattacharyya coefficient, which ranges from [0,1]. The closer it is to 1, the higher the similarity between features.

[0114] Step 5: Fuse the target template of frame t+1 and the target template of frame t through a template learning mechanism to obtain a new target template, and repeat steps 2-5;

[0115] If a static template update strategy is used, the target template will learn too much invalid feature information. Therefore, an appropriate learning rate needs to be selected to ensure the stability of long-term tracking. The tracking algorithm matches the target with the ideal feature template using the Bhattacharyya coefficient. When the target changes drastically, its value will decrease sharply, reflecting the changes in target features to some extent. The optimal target template matched in frame t+1 is γ. best (t+1), whose corresponding individual fitness value is F best If (t+1), then the adaptive learning rate is:

[0116]

[0117] Where, Φ t Φ is the learning rate for the current frame. max The peak learning rate is set to 0.04. δ and ξ are adjustment parameters that determine the fitness threshold for occlusion detection, set to 10 and 5 respectively. F best (t+1) represents the optimal fitness value of the population after the tracking algorithm completes tracking in frame t+1. The dynamic update strategy for the matching template is then:

[0118] γ(t+1)=(1-φ t+1 )·γ(t)+φ t+1·γ best (t+1)

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. An improved method for long-term vehicle target tracking optimized by a carnivorous plant algorithm, characterized in that, Includes the following steps: Step 1: From the vehicle video sequence... t Frame localization and target tracking, obtaining the first frame t The target location information of the frame is used to extract features from the tracked target to obtain the first frame. t Frame target template γ ( t ); Step 2: Based on the target template, use the improved carnivorous plant algorithm to search for several target candidate boxes; Step 3: Based on several target candidate boxes, select from the vehicle video sequence... t Several candidate templates are generated in frame +1; Step 4: Place the first t +1 frame extracted candidate template and the first t The target templates extracted from the frames are subjected to similarity matching, and the target candidate box corresponding to the candidate template with the best similarity is selected as the first frame. t The tracking target in frame +1 is used to extract features from the tracking target, resulting in the... t +1 frame target template γ best ( t+ 1); Step 5: Place the first t +1 frame target template and the first t Frame target templates are fused through a template learning mechanism to obtain new target templates. γ ( t+ 1) Repeat steps 2-5 until the last frame of the video sequence; The improved carnivorous plant algorithm's reproduction process is as follows: in, C i This refers to the plants in the carnivorous plant algorithm. P i For prey, i ≠ v , C v Carnivorous plants were randomly selected. r r The reproduction rate is set to 1.

8. f This indicates calculating the similarity between individuals. k This indicates the current iteration number of the CPA algorithm; If the population size is n The number of plants is n c The remaining individuals were all prey, numbering [number missing]. n p By setting a population partitioning mechanism, the number of prey populations adaptively decreases with the number of iterations, while the number of plants adaptively increases with the number of iterations. Therefore, a scaling factor is designed. κ Adjusting the population proportion, the number of plants is n × κ The number of prey is n ×(1- κ ); in, κ first and κ end The initial and final values ​​of the scale are set to 0.2 and 0.5 respectively. κ 1 is the damping coefficient and is greater than 0. k It is the current iteration number. k max This represents the maximum number of iterations.

2. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 1, characterized in that, The feature extraction of the tracked target specifically includes: The FHOG features and deep features of the tracked target are weighted and fused. The deep features are extracted using a ResNet-50 network model that achieved 75% success rate in ImageNet training. These deep features include Conv2-1 and Conv4-3 features from the ResNet-50 network model. Therefore, the target template... γ ( t The following is represented: in,, γ Conv2-1 , γ Conv4-3 and γ FHOG These represent the Conv2-1 features, Conv4-3 features, and FHOG features of the tracked target, respectively. λ 1. λ 2 and λ 3 represents the feature fusion coefficient.

3. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 2, characterized in that, λ 1 = 0.8 λ 2 = 0.5 and λ 3 = 0.

5.

4. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 2, characterized in that, The process of searching for several target candidate boxes based on the target template using an improved carnivorous plant algorithm specifically includes: Step 2.1: Based on the first t The target location information of the frame is obtained through the short-term memory module. t +1 frame of target location information; Step 2.2: Based on the location information of the short-term memory module, use the improved carnivorous plant algorithm to search for more target location information and obtain the location of the target candidate box; Step 2.3: Based on the improved carnivorous plant algorithm, obtain more target location information, use the scale perception factor at each target location to obtain the target scale information, and determine the size of the target candidate box.

5. The vehicle long-term target tracking method optimized by the improved carnivorous plant algorithm according to claim 4, characterized in that, The short-term memory module adopts a Gaussian distribution method and uses the position of the target in three frames of the video sequence to predict the motion trajectory. This allows the population initialization to be as close as possible to the predicted target area, making full use of the individual's search ability, shortening the global optimization time, and locating the local target faster. individual X i It follows a multidimensional Gaussian distribution, as shown in the following formula: in, p ( X i ( t+1 )) represents an individual X i The probability of occurrence, σ t+1 and μ t+1 They are respectively t+ The covariance and mean matrix of the initial population at time 1. The mean matrix is ​​related to the predicted target position, and the calculation formula is as follows: In the formula, X best for t , t -1 and t -2 frames CPA algorithm obtained the optimal solution; In the short-term memory module, the target position obtained from the previous three frames is used to determine the motion velocity in the current frame, which is then used as the mean of the Gaussian model. μ t+1 Calculate the initial position of the population in the current frame, as the first... t +1 frame contains several target location information.

6. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 1, characterized in that, The proportional sensing factor s i =[ s xi s yi ], s xi and s yi Let the length and width of the target candidate box be the scaling factors, respectively, then the new CPA individual X i =[ x i y i s xi s yi Extract FHOG features from each scale window, calculate the response values ​​corresponding to different scale factors, and select the optimal scale response to complete the scale update; Assuming the current frame tracking target window size is M O × N O The current frame image size is M P × N P Therefore, the boundary of the proportionality perception factor is: in, LB s This is the lower bound of the scale perception factor. UB s This is the upper bound of the scale perception factor. τ s The scale step size is set to 1.02; To speed up tracking, the scale is updated every 10 frames while the position is updated. Each scale update extracts... n × k max Different scale samples, n This refers to population size. k max This indicates the maximum number of iterations for the CPA algorithm.

7. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 1, characterized in that, The similarity matching in step 4 is as follows: The Bhattacharyya coefficients are used as the solution condition for template matching. The formula for the Bhattacharyya coefficients is as follows: In the formula, a and b The characteristic matrix, z The dimension of the feature matrix , ρ The obtained Bhattacharyya coefficient ranges from [0,1], with the closer it is to 1, the higher the similarity between features.

8. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 1, characterized in that, Step 5 involves fusion using a template learning mechanism, specifically as follows: Choosing an appropriate learning rate ensures the stability of long-term tracking. t+ The optimal target template matched in frame 1 is γ best ( t+ 1), the corresponding individual fitness value is F best ( t+ 1), then the first t+ The adaptive learning rate for 1 frame is: in, t+1 For the first t +1 frame learning rate, max The peak value of the learning rate. δ and ξ To adjust the parameters, F best ( t+ 1) For the tracking algorithm in the first... t+ The optimal fitness value of the population after one frame of tracking; The first t +1 frame target template and the first t The frame target templates are fused through a template learning mechanism to obtain a new target template: in, γ ( t ) is the first t Frame target template, γ best ( t+ 1) is the first t+ The target template for frame 1.

9. The method for long-term vehicle target tracking optimized by the improved carnivorous plant algorithm according to claim 8, characterized in that, max Set it to 0.

04. δ and ξ Set them to 10 and 5 respectively.

Images