Unmanned driving target recognition method and system based on variable structure convolutional neural network

Abstract

The present invention belongs to the field of unmanned driving technology, and provides an unmanned driving target recognition method and system based on a variable structure convolutional neural network. The present invention improves the backbone network of the convolutional neural network structure for the unmanned driving process, which needs to identify targets such as pedestrians, vehicles and traffic lights, defines the parameter type and value range of the search space, uses the ant colony algorithm to improve the search direction of the third defense mechanism of the crown porcupine algorithm and the conditions for triggering the fourth defense mechanism, and combines the respective advantages of the crown porcupine algorithm and the ant colony algorithm to form a fusion intelligent algorithm. The fusion intelligent algorithm that has fast convergence and good global search capability is used as a search strategy, and new structural parameter vectors are generated according to the searched space. These vectors are converted into a convolutional neural network structure model and evaluated using a fitness function, and then multiple optimizations are performed to obtain a more accurate target detection convolutional neural network model, thereby achieving more accurate and rapid target optimization.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned driving technology, and in particular relates to an unmanned driving target recognition method and system based on a variable structure convolutional neural network. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In order to ease traffic pressure and reduce traffic accidents, driverless cars have received widespread attention and are considered to be a solution that provides more efficient driving and appropriate vehicle control. There are three core technical modules in driverless technology, namely environmental perception, precise positioning, and decision-making control. Environmental perception technology is the first link of driverless driving and a prerequisite for driverless driving. Accurate and stable real-time visual perception is the premise for the safe operation of driverless vehicles. Object detection technology is the key to the pixel-level and instance-level image understanding of driverless visual perception systems.

[0004] In the field of target detection, the mainstream now is to use deep learning neural networks. The structure of neural networks is flexible and changeable, and rich experience is required to design model structures for specific tasks. When an unmanned vehicle is driving on the road, its road information, such as pedestrians, vehicles, and traffic lights, varies in size. The existing deep convolutional neural network structure is not very accurate in identifying them, especially small targets. Summary of the invention

[0005] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides an unmanned driving target recognition method and system based on a variable structure convolutional neural network, which improves the backbone network of the convolutional neural network structure to obtain a better convolutional neural network model, thereby improving the efficiency and accuracy of the target detection neural network in unmanned driving target recognition.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] The first aspect of the present invention provides an unmanned driving target recognition method based on a variable structure convolutional neural network, comprising the following steps:

[0008] Obtain unmanned driving historical image data;

[0009] The target recognition model is trained according to the acquired historical unmanned driving image data to obtain a trained target recognition model; wherein the target recognition model construction process includes:

[0010] Determine the parameter type and value range of the backbone network search space of the convolutional neural network;

[0011] The probability selection mechanism of the ant colony algorithm is introduced, and the position of the crested porcupine is continuously updated in combination with the search direction of the third defense mechanism of the crested porcupine algorithm. The condition of triggering the fourth defense mechanism based on the threshold of pheromone concentration is obtained, and the crested porcupine-ant colony combination algorithm is obtained. The crested porcupine-ant colony combination algorithm is used to search the neural network structure, optimize the parameters of the search space, and iterate the search with the prediction accuracy as the fitness function. Based on the parameters and value range of the search space corresponding to the optimal solution, the optimal target recognition model is obtained;

[0012] Based on the trained target recognition model, the target image to be detected in the unmanned driving process is recognized to obtain the target recognition result.

[0013] Furthermore, the construction process of the crown porcupine-ant colony combination algorithm includes:

[0014] In the exploration phase, the optimization is performed, and the first and second defense mechanisms of the crown porcupine algorithm are used to determine whether z(t+1), which represents the difference between the particle position after t+1 iterations and t iterations divided by the particle position at t iterations, exceeds the set error in the set number of iterations. If so, the exploration phase is jumped out and the utilization phase is executed. The odor diffusion mechanism of the crown porcupine algorithm is combined with the probability selection mechanism of the ant colony algorithm. A crown porcupine represents a solution. When a "crown porcupine" finds a promising area in the solution space, the pheromone of the area is updated, and a heuristic factor is released to guide other "crown porcupines" to move to the area.

[0015] In each iteration, each "Crown Porcupine" selects the next position direction according to the transfer probability of the ant colony, obtains the parameters used to control the search direction in the third defense of the Crown Porcupine algorithm, and then updates the position according to the third defense strategy of the Crown Porcupine, and then updates the pheromone and heuristic factor;

[0016] A threshold of pheromone concentration is set. If the pheromone concentration from node i to j is less than the threshold of pheromone concentration, the fourth defense mechanism of the crown porcupine algorithm is executed for further search. When z(t+1) exceeds the set error in the set number of iterations, the optimal solution obtained by the search is compared with the optimal solution obtained in the exploration phase, and the optimal solution is selected, thereby obtaining the crown porcupine-ant colony combination algorithm.

[0017] Furthermore, the position update formula of the crested porcupine is:

[0018]

[0019] In the formula, is the position of the i-th individual after t+1 iterations, The position of the i-th individual after t iterations, To randomly generate a binary vector containing 0 and 1, r 1 and r 2 are two random integers between [1, N], and τ 3 is a random value generated between 0 and 1, γ t As a defense factor, is the odor diffusion factor, responsible for the odor diffusion rate during the optimization process;

[0020] The parameters used to control the search direction are:

[0021]

[0022] p ij From X i Click to X j The transition probability is:

[0023]

[0024] In the formula, τ ik From X i Click to X k The pheromone concentration, w ij is the inspiration factor, α is the pheromone importance factor, β is the nutrient concentration and distance importance factor, J(i) is the set of nodes that ant k is allowed to choose in the next step, τ ij (t) is the i Click to X k The pheromone concentration, w ik is the heuristic factor from node i to node k.

[0025] Furthermore, the formula of the fourth defense mechanism is:

[0026]

[0027] In the formula, represents the position of the i-th individual after t+1 iterations, represents the optimal solution in the t-th iteration process, represents the position of the i-th individual after t iterations. At the same time, because the crested porcupine collides with the predator, It also represents the position of the predator, τ 4 , τ 5 is a random number between 0 and 1;

[0028] The average force exerted by the crested porcupine on the predator is:

[0029]

[0030] In the formula, represents the final velocity of the i-th individual at time t+1, Represents a vector of random values ​​between 0 and 1, m i represents the mass of the i-th individual or predator at the t-th iteration, represents the initial velocity of the i-th individual at time t.

[0031] Furthermore, the search parameters of the backbone network of the convolutional neural network include module type, attention mechanism type, and network width;

[0032] Among them, the module type is the basic module of the backbone network, the Bottleneck in the C3 module is replaced with GhostBottleneck, all Bottleneck modules are replaced with CrossConv modules, and ConvNextBlock is used to replace Bottleneck;

[0033] In the attention mechanism, SENet uses conventional convolution to extract features to obtain the feature map T, and then calculates the feature map T channel by channel through 1×1 convolution to obtain the channel attention of scale 1×1×C, and multiplies the channel attention with the feature map T to obtain the output feature map;

[0034] CBAM further uses the information of spatial dimension based on SENet; NAMAttention uses the same structure as CBAM and redesigns the channel attention mechanism module and the spatial attention mechanism module; TripletAttention uses three different branches to process the cross-dimensional interaction information between channel C, width W and height H.

[0035] Furthermore, the targets to be detected include pedestrians, vehicles, bushes, road traffic signs and road identification lines, and the convolutional neural network adopts YOLOvx.

[0036] Furthermore, the formula for the prediction accuracy is:

[0037] c t =0.2P+0.4R+0.1F1+0.3mAP,

[0038] Among them, mAP is the average of average precision, and its specific calculation formula is:

[0039]

[0040] In the formula, TP represents the number of correctly detected positive samples, FP represents the number of negative samples incorrectly detected as positive samples, FN represents the number of positive samples incorrectly detected as negative samples, and k is the number of categories.

[0041] A second aspect of the present invention provides an unmanned driving target recognition system based on a variable structure convolutional neural network, comprising:

[0042] The data acquisition module is configured to: acquire unmanned driving historical image data;

[0043] The target recognition model training module is configured to: train the target recognition model according to the acquired historical unmanned driving image data to obtain a trained target recognition model; wherein the target recognition model construction process includes:

[0044] Determine the parameter type and value range of the backbone network search space of the convolutional neural network;

[0045] The probability selection mechanism of the ant colony algorithm is introduced, and the position of the crowned porcupine is continuously updated in combination with the search direction of the third defense mechanism of the crowned porcupine algorithm. The condition of triggering the fourth defense mechanism based on the threshold of pheromone concentration is obtained to obtain the crowned porcupine-ant colony combination algorithm. The crowned porcupine-ant colony combination algorithm is used to search the neural network structure, and the parameters of the search space are optimized. The prediction accuracy is used as the fitness function for iterative search. The optimal target recognition model is obtained based on the parameters and value range of the search space corresponding to the optimal solution. The target recognition module is configured as follows: based on the trained target recognition model, the target image to be detected in the unmanned driving process is recognized to obtain the target recognition result.

[0046] A third aspect of the present invention provides a computer-readable storage medium.

[0047] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the unmanned driving target recognition method based on a variable structure convolutional neural network as described above.

[0048] A fourth aspect of the present invention provides a computer device.

[0049] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the unmanned driving target recognition method based on a variable structure convolutional neural network as described above are implemented.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention uses the crown porcupine-ant colony combination algorithm to optimize the neural network structure of the YOLOv5 neural backbone network to obtain a better convolutional neural network model, thereby improving the efficiency and accuracy of the target detection neural network in unmanned driving target recognition.

[0052] 2. The ant colony algorithm of the present invention improves the search direction of the third defense mechanism of the crown porcupine algorithm and the conditions for triggering the fourth defense mechanism, combines the respective advantages of the crown porcupine algorithm and the ant colony algorithm, forms a crown porcupine-ant colony combined algorithm, and achieves more accurate and rapid target optimization.

[0053] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0055] Figure 1 It is a flow chart of an unmanned driving target recognition method based on a variable structure convolutional neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0057] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0058] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0059] In view of the problem of low accuracy in target recognition in existing unmanned driving scenarios, especially in small target recognition, the present invention improves the backbone network of the target detection neural network, defines the parameter type and value range of the search space, uses the ant colony algorithm to improve the search direction of the third defense mechanism of the crown porcupine algorithm and the conditions for triggering the fourth defense mechanism, combines the respective advantages of the crown porcupine algorithm and the ant colony algorithm to form a crown porcupine-ant colony fusion algorithm, uses the obtained fusion intelligent algorithm with fast convergence and good global search capability as a search strategy, generates new structural parameter vectors according to the searched space, converts these vectors into a convolutional neural network structure model and uses a fitness function to evaluate, and then performs multiple optimizations to obtain a more accurate target detection convolutional neural network model, identifies targets based on the target detection convolutional neural network model, and improves the efficiency and accuracy of the target detection neural network in unmanned driving target recognition.

[0060] Embodiment 1

[0061] like Figure 1 As shown, this embodiment provides an unmanned driving target recognition method based on a variable structure convolutional neural network, comprising the following steps:

[0062] Step 1: Obtain unmanned driving historical image data;

[0063] In this embodiment, the target images to be detected are images of pedestrians, roads, vehicles, road signs, etc. collected in real time during the unmanned driving process.

[0064] Step 2: Train the target recognition model according to the acquired historical unmanned driving image data to obtain a trained target recognition model;

[0065] The target recognition model construction process includes:

[0066] Step 201: Determine the parameter type and value range of the backbone network search space of the convolutional neural network;

[0067] In this embodiment, the convolutional neural network uses YOLOvx, and the three types of search parameters of the backbone network are module type M, attention mechanism type A and network width W;

[0068] The module type M includes 'C3', 'C3Ghost', 'C3x', 'C3NEB';

[0069] Attention mechanism A includes 'SENet', 'CBAM', 'NAM Attention', 'TripletAttention';

[0070] The value range of network width W is 8-24.

[0071] Furthermore, in module type M, 'C3' is the basic module of the backbone network. The Bottleneck in the C3 module is replaced by GhostBottleneck, called the 'C3Ghost' module. All Bottleneck modules are replaced by CrossConv modules with lower computational overhead, called 'C3x'. ConvNextBlock is used to replace Bottleneck, called the 'C3NEB' module.

[0072] Furthermore, in the attention mechanism A, SENet uses conventional convolution for feature extraction to obtain the feature map T, and then calculates the feature map T channel by channel through 1×1 convolution to obtain the channel attention of scale 1×1×C, and multiplies the channel attention with the feature map T so that the channels with richer features get larger weights and the invalid channels get low weights, and finally the output feature map is obtained; CBAM further uses the information of spatial dimension based on SENet; NAM Attention uses the same structure as CBAM and redesigns the channel attention mechanism module and the spatial attention mechanism module; Triplet Attention uses three different branches to process the cross-dimensional interaction information between channel C, width W and height H.

[0073] Step 202: Use the ant colony algorithm to improve the search direction of the third defense mechanism of the crown porcupine algorithm and the conditions for triggering the fourth defense mechanism, and obtain the crown porcupine-ant colony combination algorithm. Use the crown porcupine-ant colony combination algorithm to search the structure of the neural network, generate new structural parameter vectors based on the searched space, convert these vectors into a convolutional neural network structure model, and use fitness to perform iterative search to obtain an optimized target recognition model.

[0074] Among them, the research design of ant colony algorithm is as follows:

[0075] Assume that at time t, the kth ant crawls from position i to the next position j, and the initial pheromone release is set to τ ij and inspiration information β ij , and set the transition probability to α is the pheromone heuristic factor, β is the expected heuristic factor, and d ij represents the distance from position i to position j, η ij is the expected heuristic function, η ij =1 / d ij , then:

[0076]

[0077] The ant colony algorithm is inspired by the foraging behavior of ants in nature. When searching for food, ants release a pheromone along the path they pass. The pheromone accumulates and evaporates along the path over time. Other ants have a higher probability of choosing a direction with a high pheromone concentration and continue to increase the pheromone concentration along the path. The pheromone concentration update formula is:

[0078]

[0079] Among them, rho is the pheromone volatility constant of the ant colony algorithm, τ mn is the pheromone concentration on the path from point m to point n, s is the serial number of the ant, q is the total number of ants, is the change in pheromone concentration on the path from point m to point n for the sth ant pair.

[0080] The first defense strategy of the crown porcupine optimization algorithm, whose race is initialized as:

[0081]

[0082] Where N represents the number of individuals, is the i-th candidate solution in the search space, are the lower and upper boundaries of the search range, respectively. is a vector randomly initialized between 0 and 1.

[0083] The first defense strategy formula of the porcupine optimization algorithm is:

[0084]

[0085] In the formula Respectively represent the position of the i-th individual after t iterations and t+1 iterations, τ 1 is a random number based on normal distribution, τ 2 is a random number between 0 and 1. represents the optimal solution in the t-iteration process, Represents the vector generated between the current crested porcupine and the crested porcupine randomly selected from the population, which is used to represent the position of the predator at the tth iteration. The specific formula is: Where r is a random number between 1 and the population size N.

[0086] In the second defense strategy of the crested porcupine algorithm, the crested porcupine uses sound methods to make noise and threaten predators. When the predator approaches the crested porcupine, the crested porcupine's voice will become louder. In order to mathematically model this behavior, the following formula is proposed:

[0087]

[0088] where r1 and r 2 are two random integers between [1, N], and τ 3 is a random value between 0 and 1, generated using random A binary vector containing 0 and 1 to cover all possible probabilities.

[0089] Equation (5) The vector in represents the position of the predator, which is between the current crested porcupine and a crested porcupine solution randomly selected from the population.

[0090] The third defense strategy of the crested porcupine algorithm is characterized by an odor strategy, where the crested porcupine secretes a foul-smelling chemical odor and spreads it around the area to prevent predators from approaching. This strategy is usually successful in resisting predators. The formula that mathematically simulates this behavior is:

[0091]

[0092] In the formula, The position of the i-th individual after t iterations, represents the position of the i-th individual after t+1 iterations, r 1 、r 2 、r 3 is a random number between [1, N], N is the population size, t is the number of current iterations, τ 3 Is a random number between 0 and 1. It is a parameter used to control the search direction, and its definition equation is:

[0093]

[0094] is a vector containing randomly generated values ​​between 0 and 1, γ t is the defense factor, and its definition equation is:

[0095]

[0096] Where, t max is the maximum number of iterations, is the odor diffusion factor, which is responsible for the odor diffusion rate during the optimization process. In this embodiment, the various values ​​of this factor are between 0.3 and 2.6.

[0097] If the values ​​of this factor are small, the diffusion rate is very small, so solutions close to random selection are exploited, on the contrary, if these values ​​are high, the odor diffusion rate is high, so solutions far from random selection are exploited. Its definition equation is:

[0098]

[0099] In the formula represents the objective function, and ε represents a small value to avoid division by zero. Used to simulate three possible situations in this strategy:

[0100] (1) When When it is equal to 0, the crested porcupine will stop diffusing its scent because the predator will stop moving due to fear of the crested porcupine, so the distance between the predator and the crested porcupine remains unchanged;

[0101] (2) When When it is equal to 1, the crested porcupine will emit a significant scent due to the presence of a predator nearby;

[0102] (3) When Between 0 and 1, predators keep a safe distance from crested porcupines, so there is no need to significantly release their scent.

[0103] The fourth defense strategy of the crested porcupine algorithm is characterized by a physical behavior. When the predator is very close, the crested porcupine hits the predator with a physical attack. During the physical attack, the two bodies are tightly pressed together, just like an inelastic collision in one-dimensional space. The formula is:

[0104]

[0105] In the formula, represents the position of the i-th individual after t+1 iterations, represents the optimal solution in the t-th iteration process, represents the position of the i-th individual after t iterations. At the same time, because the crested porcupine collides with the predator, It also represents the position of the predator, τ 4 , τ 5 Is a random number between 0 and 1.

[0106] It represents the average force exerted by the crested porcupine on the predator. The specific formula is:

[0107]

[0108] In the formula, represents the final velocity of the ith individual at time t+1 and is assigned according to a solution randomly selected from the current population, so represents the initial velocity of the ith individual at time t, Represents a vector of random values ​​between 0 and 1, m i It represents the mass of the i-th individual (predator) at the t-th iteration. The specific formula is: In the formula, f(·) represents the objective function, and ε represents a small value to avoid division by zero.

[0109] δ is a parameter used to control the search direction and is defined using the formula defined as is a vector containing randomly generated values between 0 and 1;

[0110] γ t is the defense factor and is represented by the formula where rand represents a random number between 0 and 1, t is the current iteration number, and t max is the maximum number of iterations; α is the convergence speed factor, which determines the convergence speed of the algorithm. After testing, the best effect is achieved when α = 0.1.

[0111]

[0112] In the Crown Porcupine algorithm, the Crown Porcupine algorithm is based on the designed exploration and exploitation phases, and each phase has two defense strategies for the Crown Porcupine. Among them, the exploration phase includes the line of sight and sound defense strategies, and the exploitation phase includes the smell and physical attack defense strategies. In the exploration phase, when τ6 and τ7 are two randomly generated values between 0 and 1, when τ6 < τ7, a trade-off between the line of sight defense strategy and the sound defense strategy will be executed. In the exploitation phase, the smell defense strategy is executed when τ10 < Tf; otherwise, the physical attack defense strategy will be executed.

[0113] Tf is a predetermined constant value between 0 and 1, which is used to make a trade-off between local exploitation (the third defense mechanism) and global exploitation (the fourth defense mechanism) in the case of the combination of the third and fourth defense mechanisms; a small value of this factor may lead to stagnation at the local minimum, while a high value may lead to a slow convergence speed.

[0114] In this embodiment, through a large number of experiments, the optimal value T of this parameter is obtained f to be 0.9.

[0115] First, the exploration phase is carried out for optimization. In this phase, the first and second defense mechanisms of the Crown Porcupine algorithm are adopted, and let It represents the difference between the particle position after iteration t+1 and iteration t divided by the particle position at iteration t. When z(t+1)<1% appears 30 times in a row during the optimization process of the crown porcupine algorithm, it means that the particle position changes very little, and the first stage is jumped out, and the second stage is executed. The characteristics of the ant colony algorithm are integrated into the third defense mechanism of the crown porcupine algorithm. The odor diffusion mechanism of the crown porcupine algorithm can be combined with the probability selection mechanism of the ant colony algorithm. When a "crown porcupine" (representing a solution) finds a promising area in the solution space, it not only updates the pheromone in the area, but also releases a heuristic factor to guide other "crown porcupines" to move to the area.

[0116] When selecting the next node, combining pheromone and heuristic factors, the formula is:

[0117]

[0118] Among them, p ij From X i Click to X j The transition probability, τ ik From X i Click to X k The pheromone concentration, w ij is the inspiration factor, α is the pheromone importance factor, β is the nutrient concentration and distance importance factor. J(i) is the set of nodes that ant k is allowed to choose in the next step, τ ij (t) is the i Click to X k The pheromone concentration, w ik is the heuristic factor from node i to node k.

[0119] In each iteration, each "Crown Porcupine" selects the next position direction according to the above probability formula. In the third defense of the Crown Porcupine algorithm, δ is a parameter used to control the search direction. Combining it with the ant colony algorithm, the improved formula is:

[0120]

[0121] Then the position is updated according to the third defense strategy of the crested porcupine, the formula is:

[0122]

[0123] Pheromones and heuristic factors are updated accordingly.

[0124] Set a threshold value of pheromone concentration τ threshold , if the pheromone concentration τ from node i to j ij If it is greater than or equal to this threshold, the pheromone concentration in the area is considered to be high.

[0125] If it is located in an area with high pheromone concentration, τ ij Less than τ threshold , the fourth defense mechanism of the crown porcupine algorithm will be executed, the formula is:

[0126]

[0127] That is, a more intensive search is performed in the area.

[0128] When z(t+1)<1% occurs for 30 consecutive times, compare the optimal solution obtained by searching with the optimal solution obtained in the first stage of the crown porcupine algorithm, select the best one, and use M n Record the number of times. n The iteration stops when ≥5 or the number of iterations reaches the maximum number of iterations. Thus, a fused intelligent algorithm is obtained.

[0129] The crown porcupine-ant colony algorithm is used to search for neural network structures and determine the parameter type and value range of the search space. The crown porcupine-ant colony algorithm is used to optimize the three determined search parameters. t =0.2P+0.4R+0.1F1+0.3mAP The predicted accuracy is used as the fitness function, and an iterative search is performed to obtain the optimized target detection convolutional neural network model.

[0130] Step 3: Based on the trained target recognition model, the target image to be detected during the unmanned driving process is recognized to obtain the target recognition result.

[0131] Embodiment 2

[0132] This embodiment provides an unmanned driving target recognition system based on a variable structure convolutional neural network, including:

[0133] The data acquisition module is configured to: acquire unmanned driving historical image data;

[0134] The target recognition model training module is configured to: train the target recognition model according to the acquired historical unmanned driving image data to obtain a trained target recognition model; wherein the target recognition model construction process includes:

[0135] Determine the parameter type and value range of the backbone network search space of the convolutional neural network;

[0136] The probability selection mechanism of the ant colony algorithm is introduced, and the position of the crested porcupine is continuously updated in combination with the search direction of the third defense mechanism of the crested porcupine algorithm. The condition of triggering the fourth defense mechanism based on the threshold of pheromone concentration is obtained, and the crested porcupine-ant colony combination algorithm is obtained. The crested porcupine-ant colony combination algorithm is used to search the neural network structure, optimize the parameters of the search space, and iterate the search with the prediction accuracy as the fitness function. Based on the parameters and value range of the search space corresponding to the optimal solution, the optimal target recognition model is obtained;

[0137] The target recognition module is configured to: recognize the target image to be detected during the unmanned driving process based on the trained target recognition model to obtain a target recognition result.

[0138] Embodiment 3

[0139] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the unmanned driving target recognition method based on a variable structure convolutional neural network as described above are implemented.

[0140] Embodiment 4

[0141] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the unmanned driving target recognition method based on a variable structure convolutional neural network as described above are implemented.

[0142] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0143] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0144] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0146] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0147] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An unmanned driving target recognition method based on a variable structure convolutional neural network is characterized in that: The steps include: Obtain unmanned driving historical image data; The target recognition model is trained according to the acquired historical unmanned driving image data to obtain a trained target recognition model; wherein the target recognition model construction process includes: Determine the parameter type and value range of the backbone network search space of the convolutional neural network; the parameter type of the backbone network search space of the convolutional neural network includes module type, attention mechanism type and network width; The probability selection mechanism of the ant colony algorithm is introduced, and the position of the crested porcupine is continuously updated in combination with the search direction of the third defense mechanism of the crested porcupine algorithm. The condition of triggering the fourth defense mechanism based on the threshold of pheromone concentration is obtained, and the crested porcupine-ant colony combination algorithm is obtained. The crested porcupine-ant colony combination algorithm is used to search the neural network structure, optimize the parameters of the search space, and iterate the search with the prediction accuracy as the fitness function. Based on the parameters and value range of the search space corresponding to the optimal solution, the optimal target recognition model is obtained; The construction process of the crown porcupine-ant colony combination algorithm includes: In the exploration phase, the optimization is performed, and the first and second defense mechanisms of the crested porcupine algorithm are used to determine whether z(t+1), which represents the difference between the particle position after t+1 iterations and t iterations divided by the particle position at t iterations, exceeds the set error for consecutively set iteration times. If so, the exploration phase is jumped out and the utilization phase is executed. The odor diffusion mechanism of the crested porcupine algorithm is combined with the probability selection mechanism of the ant colony algorithm. A crested porcupine represents a solution. When a crested porcupine finds a promising area in the solution space, the pheromone of the area is updated, and a heuristic factor is released to guide other crested porcupines to move to the area. In each iteration, each crested porcupine selects the next position direction according to the transfer probability of the ant colony, obtains the parameters used to control the search direction in the third defense of the crested porcupine algorithm, and then updates the position according to the third defense strategy of the crested porcupine, and then updates the pheromone and heuristic factor; Set a threshold of pheromone concentration. If the pheromone concentration from node i to j is less than the threshold of pheromone concentration, execute the fourth defense mechanism of the crown porcupine algorithm to further search. When z(t+1) exceeds the set error in the set number of iterations, compare the optimal solution obtained by the search with the optimal solution obtained in the exploration phase, and select the optimal solution, thus obtaining the crown porcupine-ant colony combination algorithm. Based on the trained target recognition model, the target image to be detected in the unmanned driving process is recognized to obtain the target recognition result.

2. The unmanned driving target recognition method based on variable structure convolutional neural network according to claim 1, characterized in that: The position update formula of the crested porcupine is: In the formula, is the position of the i-th individual after t+1 iterations, The position of the i-th individual after t iterations, To randomly generate a binary vector containing 0 and 1, r1, r2, and r3 are random integers between [1, N], and τ3 is a random value generated between 0 and 1, γ t As a defense factor, is the odor diffusion factor, responsible for the odor diffusion rate during the optimization process; The parameters used to control the search direction are: p ij From X i Click to X j The transition probability is: In the formula, τ ij () is from X i Click to X j The pheromone concentration at the point, w ij () is the inspiration factor, α is the pheromone importance factor, β is the nutrient concentration and distance importance factor, J(i) is the node set that ant k is allowed to choose in the next step, τ ik From X i Click to X k The pheromone concentration at the point, w ik is the heuristic factor from node i to node k.

3. The unmanned driving target recognition method based on variable structure convolutional neural network as claimed in claim 2, It is characterized in that The formula for the fourth defense mechanism is: In the formula, represents the position of the i-th individual after t+1 iterations, represents the optimal solution in the t-iteration process, represents the position of the i-th individual after t iterations. At the same time, because the crested porcupine collides with the predator, It also represents the position of the predator, τ4 and τ5 are random numbers between 0 and 1; γ t It is a defense factor; The average force exerted by the crested porcupine on the predator is: In the formula, represents the final velocity of the i-th individual at time t+1, Represents a vector of random values ​​between 0 and 1, m i represents the mass of the i-th individual or predator at the t-th iteration, represents the initial velocity of the i-th individual at time t.

4. The unmanned driving target recognition method based on variable structure convolutional neural network according to claim 1, characterized in that: The module type is the basic module of the backbone network. The Bottleneck in the C3 module is replaced with GhostBottleneck, all Bottleneck modules are replaced with CrossConv modules, and ConvNextBlock is used to replace Bottleneck. In the attention mechanism, SENet uses convolution operation to extract features to obtain feature map T, and then calculates the feature map T channel by channel through 1×1 convolution to obtain channel attention with a scale of 1×1×C, and multiplies the channel attention with the feature map T to obtain the output feature map; CBAM further uses the information of spatial dimension based on SENet; NAMAttention uses the same structure as CBAM and redesigns the channel attention mechanism module and the spatial attention mechanism module; TripletAttention uses three different branches to process the cross-dimensional interaction information between channel C, width W and height H.

5. The unmanned driving target recognition method based on variable structure convolutional neural network according to claim 1, characterized in that: The targets to be detected include pedestrians, vehicles, bushes, road traffic signs and road identification lines, and the convolutional neural network adopts YOLOvx.

6. The unmanned driving target recognition method based on variable structure convolutional neural network according to claim 1, characterized in that: The formula for the prediction accuracy is: c t =0.2P+0.4R+0.1F1+0.3mAP, Among them, mAP is the average of average precision, and its specific calculation formula is as follows: In the formula, TP represents the number of correctly detected positive samples, FP represents the number of negative samples incorrectly detected as positive samples, FN represents the number of positive samples incorrectly detected as negative samples, and k is the number of categories.

7. Unmanned driving target recognition system based on variable structure convolutional neural network, characterized in that: include: The data acquisition module is configured to: acquire unmanned driving historical image data; The target recognition model training module is configured to: train the target recognition model according to the acquired historical unmanned driving image data to obtain a trained target recognition model; wherein the target recognition model construction process includes: Determine the parameter type and value range of the backbone network search space of the convolutional neural network; the parameter type of the backbone network search space of the convolutional neural network includes module type, attention mechanism type and network width; The probability selection mechanism of the ant colony algorithm is introduced, and the position of the crested porcupine is continuously updated in combination with the search direction of the third defense mechanism of the crested porcupine algorithm. The condition of triggering the fourth defense mechanism based on the threshold of pheromone concentration is obtained, and the crested porcupine-ant colony combination algorithm is obtained. The crested porcupine-ant colony combination algorithm is used to search the neural network structure, optimize the parameters of the search space, and iterate the search with the prediction accuracy as the fitness function. Based on the parameters and value range of the search space corresponding to the optimal solution, the optimal target recognition model is obtained; The construction process of the crown porcupine-ant colony combination algorithm includes: In the exploration phase, the optimization is performed, and the first and second defense mechanisms of the crested porcupine algorithm are used to determine whether z(t+1), which represents the difference between the particle position after t+1 iterations and t iterations divided by the particle position at t iterations, exceeds the set error for consecutively set iteration times. If so, the exploration phase is jumped out and the utilization phase is executed. The odor diffusion mechanism of the crested porcupine algorithm is combined with the probability selection mechanism of the ant colony algorithm. A crested porcupine represents a solution. When a crested porcupine finds a promising area in the solution space, the pheromone of the area is updated, and a heuristic factor is released to guide other crested porcupines to move to the area. In each iteration, each crested porcupine selects the next position direction according to the transfer probability of the ant colony, obtains the parameters used to control the search direction in the third defense of the crested porcupine algorithm, and then updates the position according to the third defense strategy of the crested porcupine, and then updates the pheromone and heuristic factor; Set a threshold of pheromone concentration. If the pheromone concentration from node i to j is less than the threshold of pheromone concentration, execute the fourth defense mechanism of the crown porcupine algorithm to further search. When z(t+1) exceeds the set error in the set number of iterations, compare the optimal solution obtained by the search with the optimal solution obtained in the exploration phase, and select the optimal solution, thus obtaining the crown porcupine-ant colony combination algorithm. The target recognition module is configured to: recognize the target image to be detected during the unmanned driving process based on the trained target recognition model to obtain a target recognition result.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the unmanned driving target recognition method based on a variable structure convolutional neural network as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the unmanned driving target recognition method based on a variable structure convolutional neural network as described in any one of claims 1 to 6 are implemented.

Images