Front vehicle start-stop detection method, system and equipment based on monocular camera

Abstract

The invention provides a front vehicle start-stop detection method, system and device based on a monocular camera, the monocular camera on a vehicle collects real-time driving images in front of the vehicle, and image pairs are generated through a time sequence construction technology. An image pair is input into a MobileNet neural network model, and feature vectors of different levels are extracted through three times of down-sampling. The extracted feature vectors are input into a classification network model, and the model performs state classification according to the image content. The classification result and the feature vector are jointly input into a depth calculation model, and the model generates a depth information value reflecting the distance of the front vehicle through layer-by-layer convolution calculation and up-sampling. According to the depth information value, in combination with a preset safety distance threshold value, safety distance signals of different levels are generated, and a driver is reminded through a vehicle displayer. A monocular camera is used for replacing a binocular camera or a radar, and system installation and maintenance are simplified. And a lightweight MobileNet neural network model and a depth calculation model are adopted, so that the calculation power demand is remarkably reduced, and the scheme is adaptive to a traditional vehicle model.

Description

Technical Field

[0001] The present invention relates to the field of automobile technology, and in particular to a method, system and device for detecting start and stop of a preceding vehicle based on a monocular camera. Background Art

[0002] As cars become more intelligent, they have more and more functions to assist drivers in driving, such as adaptive cruise control, lane assistance, and start-stop detection of the vehicle in front.

[0003] The start-stop detection of the front vehicle is mainly to help the driver drive better in complex traffic environments, especially in traffic congestion and following vehicles. It can promptly remind the driver of the start-stop status of the front vehicle, so as to maintain a safe distance from the front vehicle.

[0004] The existing front vehicle start-stop detection mainly relies on various sensors to obtain the surrounding environment information of the car, and then calculates the surrounding environment information to judge the calculation results. Finally, the system makes a result to prompt the driver to assist the driver in driving the car.

[0005] The existing sensors for obtaining information about the surrounding environment of a car are divided into two types: one is to obtain information by radar, and the other is to obtain information by binocular camera.

[0006] For radar sensors, since radar sensors use reflected waves to identify whether the vehicle in front is started or stopped, the detection accuracy for pedestrians in traffic jams is not as good as that of binocular cameras. Secondly, radar sensors cannot obtain the color and shape of objects and can only judge the surrounding environment information through reflected waves. At the same time, radar sensors (lidar sensors) are also affected by the reflectivity of the surface of objects during use, especially on the surface of smooth or high-reflectivity objects, which will result in insufficient reflected waves, making it difficult for radar sensors to detect, and thus unable to remind the driver in time.

[0007] For binocular cameras, when detecting the start and stop of the front vehicle, a large number of images need to be obtained through the binocular cameras, and then image processing algorithms are used to calculate the parallax and infer the depth information from it. This requires high computing resources, so the computing power requirements of the car in the process of detecting the start and stop of the front vehicle are very large. Secondly, during use, as the car drives on the road, it will produce bumps, shakes and ups and downs. In the case of long-term use, the binocular camera will produce displacement. At this time, the two cameras need to be calibrated regularly to ensure the relative position and direction of the two cameras are accurate.

[0008] Secondly, for some traditional car models, since there are no cameras, two cameras need to be installed at the front of the car to have the front vehicle start-stop detection function. However, for binocular cameras, the required computing power is very high. The computing power of traditional car models is relatively low, making it difficult to apply to traditional car models. Summary of the invention

[0009] Based on this, it is necessary to propose a method, system and equipment for detecting the start and stop of the preceding vehicle based on a monocular camera to address the above-mentioned problems.

[0010] A method for detecting a start or stop of a preceding vehicle based on a monocular camera, the method comprising:

[0011] Acquire driving images of the vehicle during driving by using a monocular camera in front of the vehicle, and construct image pairs for the driving images according to a time sequence;

[0012] Inputting the image pair into a MobileNet neural network model, and the MobileNet neural network model inputting the image pair into a feature extraction module to obtain a feature extraction vector;

[0013] Inputting the feature extraction vector into a classification network model, the classification network model classifies the feature extraction vector to obtain a classification result;

[0014] Inputting the classification result and the feature extraction vector into a depth calculation model to calculate the depth information value;

[0015] A safety distance signal is generated according to the depth information value, and the safety distance signal is displayed on a vehicle display to remind the driver.

[0016] In at least one embodiment of the present application, the step of inputting the image pair into the MobileNet neural network model, the MobileNet neural network model inputting the image pair into the feature extraction module, and obtaining the feature extraction vector comprises:

[0017] The feature extraction module downsamples the image pair three times to obtain feature extraction vectors after the three downsamplings, wherein the feature extraction vectors include: a first feature vector of the first downsampling, a second feature vector of the second downsampling, and a third feature vector of the third downsampling.

[0018] In at least one embodiment of the present application, the step of inputting the feature extraction vector into a classification network model, the classification network model classifying the feature extraction vector, and obtaining the classification result comprises:

[0019] Analyzing the feature extraction vector to obtain a second feature vector;

[0020] The second feature vector is input into a classification network model, and the classification network model classifies the second feature vector to obtain a classification result, wherein the classification result includes: a start forward state, a deceleration stop state, and a no state result category.

[0021] In at least one embodiment of the present application, the step of inputting the feature extraction vector into a classification network model, the classification network model classifying the feature extraction vector, and obtaining the classification result includes:

[0022] It is determined whether the classification result is stateless. If the classification result is stateless, the image processing of the next cycle is performed.

[0023] In at least one embodiment of the present application, the method further includes:

[0024] If the classification result is determined to be a start-up forward state or a deceleration stop state, a first prompt signal is generated according to the state corresponding to the classification result, and the first prompt signal is displayed on a vehicle display.

[0025] In at least one embodiment of the present application, the method further includes:

[0026] If the classification result is determined to be a start-up forward state or a deceleration stop state, the feature extraction vector is input into a depth calculation model;

[0027] The deep computing model analyzes the feature extraction vector to obtain a first feature vector, a second feature vector, and a third feature vector;

[0028] The depth calculation model performs convolution calculation and upsampling on the third feature vector to obtain a first depth feature;

[0029] The depth calculation model performs channel superposition on the second feature vector and the first depth feature and then performs convolution calculation and upsampling to obtain the second depth feature;

[0030] The depth calculation model performs channel superposition on the first feature vector and the second depth feature and then performs convolution calculation and upsampling to obtain a depth feature vector;

[0031] The depth calculation model uses a 1x1 convolution kernel to transform the depth feature vector to obtain a depth information value.

[0032] In at least one embodiment of the present application, the step of generating a safety distance signal according to the depth information value and displaying the safety distance signal on a vehicle display to remind the driver includes:

[0033] The distance value L to the preceding vehicle is calculated based on the depth information value d:

[0034] L = αd, where α is a constant;

[0035] If the distance value L is less than or equal to the first preset safety distance value, a first distance reminder signal is generated;

[0036] If the first preset safety distance value<distance value≤second preset safety distance value, a second distance reminder signal is generated.

[0037] In at least one embodiment of the present application, the specific steps of acquiring a driving image of the vehicle during driving by a monocular camera in front of the vehicle and constructing an image pair based on the time sequence of the driving image include:

[0038] Acquire a current frame image and a fourth frame image before the current frame image in the driving image, and superimpose the current frame image and the fourth frame image on a channel to obtain a superimposed image pair;

[0039] Performing image scaling and image normalization processing on the image pair to obtain a processed image pair;

[0040] The processed image pairs are input into the MobileNet neural network model.

[0041] A front vehicle start-stop detection system based on a monocular camera is applied to the front vehicle start-stop detection method based on a monocular camera as described above, and the system comprises:

[0042] An image acquisition module, used to acquire images from a monocular camera in front of the vehicle during driving;

[0043] MobileNet neural network model, performs feature extraction on image pairs and obtains feature extraction vectors;

[0044] The classification network model classifies the input feature extraction vector to obtain the classification result;

[0045] The depth calculation model calculates the feature extraction vector according to the classification result to obtain the calculated depth information value;

[0046] A safety distance generation module generates a safety distance signal according to the depth information value;

[0047] The monocular camera-based front vehicle start-stop detection system performs the following steps:

[0048] Acquire driving images of the vehicle during driving by using a monocular camera in front of the vehicle, and construct image pairs for the driving images according to a time sequence;

[0049] Inputting the image pair into a MobileNet neural network model, and the MobileNet neural network model inputting the image pair into a feature extraction module to obtain a feature extraction vector;

[0050] Inputting the feature extraction vector into a classification network model, the classification network model classifies the feature extraction vector to obtain a classification result;

[0051] Inputting the classification result and the feature extraction vector into a depth calculation model to calculate the depth information value;

[0052] A safety distance signal is generated according to the depth information value, and the safety distance signal is displayed on a vehicle display to remind the driver.

[0053] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0054] Acquire driving images of the vehicle during driving by using a monocular camera in front of the vehicle, and construct image pairs for the driving images according to a time sequence;

[0055] Inputting the image pair into a MobileNet neural network model, and the MobileNet neural network model inputting the image pair into a feature extraction module to obtain a feature extraction vector;

[0056] Inputting the feature extraction vector into a classification network model, the classification network model classifies the feature extraction vector to obtain a classification result;

[0057] Inputting the classification result and the feature extraction vector into a depth calculation model to calculate the depth information value;

[0058] A safety distance signal is generated according to the depth information value, and the safety distance signal is displayed on a vehicle display to remind the driver.

[0059] The implementation of the monocular camera-based front vehicle start / stop detection method, system and device of the present invention will have at least the following beneficial effects:

[0060] The present invention proposes a front vehicle start-stop detection method, system and device based on a monocular camera. The monocular camera on the vehicle collects real-time driving images in front of the vehicle, generates image pairs through time series construction technology, and ensures that the input data has depth information potential.

[0061] The image pair is input into the MobileNet neural network model and downsampled three times to extract feature vectors of different levels. These features retain the key information of the front vehicle’s motion state and scene.

[0062] The extracted feature vector is input into the classification network model, and the model classifies the state according to the image content to determine whether the front vehicle is started, stopped, or in an idle state.

[0063] The classification result and the feature vector are input into the depth calculation model together. The model generates a depth information value reflecting the distance of the vehicle in front through layer-by-layer convolution calculation and upsampling.

[0064] According to the depth information value and the preset safety distance threshold, different levels of safety distance signals are generated to remind the driver through the vehicle display.

[0065] Using a monocular camera instead of a binocular camera or radar can significantly reduce hardware costs and simplify system installation and maintenance.

[0066] The lightweight MobileNet neural network model and optimized deep computing model are used to significantly reduce computing power requirements, making the solution suitable for traditional models.

[0067] By linking image pairs, classification network models, and deep computing models in a time-series manner, the accuracy and reliability of preceding vehicle start / stop detection are improved.

[0068] The detection solution for complex traffic scenarios such as congestion and following vehicles provides drivers with accurate safety distance information prompts.

[0069] Generate and display safe distance signals in a timely manner, significantly improving the driver's reaction speed in complex road conditions and reducing the probability of accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0071] in:

[0072] Figure 1 is a flow chart of a method for detecting start and stop of a preceding vehicle based on a monocular camera in one embodiment;

[0073] Figure 2 is a flow chart of a method for detecting start and stop of a preceding vehicle based on a monocular camera in another embodiment;

[0074] Figure 3 This is a flow chart of a method for detecting start and stop of a preceding vehicle based on a monocular camera in yet another embodiment;

[0075] Figure 4 is a structural block diagram of a front vehicle start / stop detection system based on a monocular camera in one embodiment;

[0076] Figure 5 FIG. 4 is a structural block diagram of a computer device in one embodiment.

[0077] 100. Front vehicle start / stop detection system based on monocular camera;

[0078] 110. Image acquisition module; 120. MobileNet neural network model; 130. Classification network model; 140. Depth calculation model; 150. Safety distance generation module. DETAILED DESCRIPTION

[0079] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0080] A method for detecting a start or stop of a preceding vehicle based on a monocular camera, the method comprising:

[0081] S101, obtaining a driving image of the vehicle during driving through a monocular camera in front of the vehicle, and constructing an image pair for the driving image according to a time sequence;

[0082] S102, inputting the image pair into a MobileNet neural network model, and the MobileNet neural network model inputting the image pair into a feature extraction module to obtain a feature extraction vector;

[0083] S103, inputting the feature extraction vector into a classification network model, and the classification network model classifies the feature extraction vector to obtain a classification result;

[0084] S104, inputting the classification result and the feature extraction vector into a depth calculation model to calculate a depth information value;

[0085] S105 , generating a safety distance signal according to the depth information value, and displaying the safety distance signal on a vehicle display to remind the driver.

[0086] Please refer to Figure 1 ,In this implementation, a monocular camera on the vehicle collects real-time driving images in front of the vehicle, and generates ,image pairs through the time sequence construction technique to ensure that the input ,data has the potential of depth information.

[0087] The image pair is input into the MobileNet neural network model, and after three downsamplings, feature vectors of different levels are extracted. These features retain the key information of the front vehicle's motion state and scene. The extracted feature vectors are input into the classification network model, and the model classifies the state according to the image content to determine whether the front vehicle is started, stopped, or in an idle state.

[0088] The classification result and the feature vector are input into the depth calculation model. The model generates a depth information value reflecting the distance of the vehicle in front through layer-by-layer convolution calculation and upsampling. According to the depth information value, combined with the preset safety distance threshold, different levels of safety distance signals are generated to remind the driver through the vehicle display.

[0089] Using a monocular camera instead of a binocular camera or radar can significantly reduce hardware costs and simplify system installation and maintenance. Using a lightweight MobileNet neural network model and an optimized deep computing model can significantly reduce computing power requirements, making the solution suitable for traditional models.

[0090] The accuracy and reliability of the start-stop detection of the preceding vehicle are improved by linking the image pairs, classification network models, and deep computing models in a time-series manner. The detection scheme for complex traffic scenarios such as congestion and following vehicles provides drivers with accurate safety distance information prompts. The safety distance signal is generated and displayed in a timely manner, which significantly improves the driver's reaction speed in complex road conditions and reduces the probability of accidents.

[0091] It should be noted that a monocular camera is used to obtain image information in front of the vehicle, rather than a traditional binocular camera or radar. This reduces hardware dependence and cost. The time-series component image pair is composed of a combination of different frame images in time, laying the foundation for subsequent depth information derivation.

[0092] Without increasing the hardware complexity, the monocular camera captures continuous image sequences, solving the limitation that traditional single-frame images cannot effectively infer depth information. The hardware requirements are reduced, and the information input form is optimized through the design of time-series image pairs, providing a more stable data source for depth and state detection.

[0093] Use the MobileNet neural network model for feature extraction. MobileNet is designed to be lightweight and is suitable for automotive electronic control units (ECUs) with limited computing resources. Extract key image features related to the start and stop of the preceding vehicle to reduce the computing burden while ensuring the accuracy of the extraction results. Provides an efficient and low-computing feature extraction method that is suitable for traditional models with limited resources.

[0094] The classification network model determines the vehicle state based on the feature extraction vector, including "start forward state", "deceleration stop state" and "no state". It realizes efficient judgment of the movement state of the front vehicle and assists the driver in decision-making. Accurate classification can effectively reduce false alarms and improve the reliability of the system. At the same time, the classification accuracy is directly related to driving safety. The depth calculation model uses the feature vector and classification results as input, and combines convolution calculation, upsampling and other operations to derive depth information. Under the limitation of the monocular camera, the depth information of the front vehicle is obtained through calculation and analysis methods.

[0095] It makes up for the defect that monocular cameras cannot directly obtain depth and provides necessary data for safe distance calculation.

[0096] Calculate the distance to the vehicle ahead based on the depth information value and determine whether it exceeds the preset safety range. Prompt the driver by generating different levels of warning signals. Provide driving assistance functions to remind the driver to take timely measures to ensure safety. Improve the driving experience in congestion and following car scenarios and reduce the risk of accidents.

[0097] It should be further explained that the monocular camera can be a vehicle-mounted camera, such as a camera installed in a hole in the front of the vehicle, or a camera of a driving recorder.

[0098] In at least one embodiment of the present application, the step of inputting the image pair into the MobileNet neural network model, the MobileNet neural network model inputting the image pair into the feature extraction module, and obtaining the feature extraction vector comprises:

[0099] S204, the feature extraction module downsamples the image pair three times to obtain feature extraction vectors after the three downsamplings, wherein the feature extraction vectors include: a first feature vector of the first downsampling, a second feature vector of the second downsampling, and a third feature vector of the third downsampling.

[0100] Please refer to Figure 1-Figure 3 In this embodiment, the image pairs are constructed through time sequence (such as superposition of current frame and historical frame) to form input data, and are pre-processed (normalization, scaling, etc.).

[0101] The image pair is input into the feature extraction module of the MobileNet model and the convolution calculation begins.

[0102] First downsampling: extract edges and simple geometric features in the image and generate the first feature vector.

[0103] Second downsampling: Through further convolution, more complex shape features and texture features are captured to generate the second feature vector.

[0104] The third downsampling: further abstraction, extracting global high-level semantic features, and generating the third feature vector.

[0105] After three downsamplings, feature extraction vectors containing multi-level information are generated, which provide basic data for subsequent classification and depth calculation. The three-downsampling method is used to convert complex image inputs into multi-level feature vectors, which not only improves processing efficiency but also enhances the model's ability to express key information.

[0106] The lightweight design of MobileNet significantly reduces the computational burden of the model, making it suitable for embedded platforms or traditional models with limited computing power. By extracting features in layers, the multi-scale information of the image from local to global is retained, providing more comprehensive data support for classification and depth inference. The feature extraction process is highly robust and can adapt to camera jitter and changes in image quality, while providing data support for the functional expansion of subsequent modules (such as more complex state classification or environmental modeling).

[0107] It should be noted that the input image pairs are processed using the MobileNet model to extract feature information from the images. The MobileNet model has low computational complexity and can be operated in real time on low-computing devices, making it particularly suitable for embedded systems in the automotive field. The image pairs are fed into the feature extraction module inside the MobileNet model, which extracts key information from the image through convolution operations and converts it into a high-dimensional feature vector. The original image pairs are converted into feature representations that can be processed by subsequent classification networks and deep computing networks.

[0108] By extracting feature vectors, the redundancy of the original image information is greatly reduced, while retaining key visual features, optimizing the efficiency and accuracy of subsequent model processing. The feature extraction module uses downsampling techniques (such as deep separable convolution) to perform three dimensionality reduction operations on the input image. Each downsampling reduces the spatial resolution of the feature map, but enhances its semantic expression ability. The low-level, mid-level and high-level features of the input image are extracted hierarchically, and a multi-scale, multi-level feature expression is gradually constructed. Hierarchical feature extraction allows the model to capture local detail features and perceive global semantic information, laying the foundation for subsequent deep calculations and classification tasks.

[0109] Each downsampling generates a feature vector:

[0110] The first eigenvector of the first downsampling: captures the basic features of the input image pair (such as edges, textures, etc.). The second eigenvector of the second downsampling: extracts mid-level features (such as object shape, local patterns). The third eigenvector of the third downsampling: obtains high-level semantic features (such as scene understanding and target motion status).

[0111] Through multi-level feature extraction, the information of the input image pair is fully described, which facilitates the subsequent processing modules to perform deep calculations and state classification. It provides multi-scale information, making the model more robust when processing complex scenes, and can take into account both fine details and global target information. The image pair is formed into input data through time series construction (such as superposition of the current frame and the historical frame), and preprocessed (normalization, scaling, etc.). The image pair is input to the feature extraction module of the MobileNet model, and convolution calculations begin.

[0112] First downsampling: extract edges and simple geometric features in the image and generate the first feature vector.

[0113] Second downsampling: Through further convolution, more complex shape features and texture features are captured to generate the second feature vector.

[0114] The third downsampling: further abstraction, extracting global high-level semantic features, and generating the third feature vector.

[0115] After three downsamplings, feature extraction vectors containing multi-level information are generated, which provide basic data for subsequent classification and depth calculation. The three-downsampling method is used to convert complex image inputs into multi-level feature vectors, which not only improves processing efficiency but also enhances the model's ability to express key information.

[0116] The lightweight design of MobileNet significantly reduces the computational burden of the model, making it suitable for embedded platforms or traditional models with limited computing power. By extracting features in layers, the multi-scale information of the image from local to global is retained, providing more comprehensive data support for classification and depth inference. The feature extraction process is highly robust and can adapt to camera jitter and changes in image quality, while providing data support for the functional expansion of subsequent modules (such as more complex state classification or environmental modeling).

[0117] In at least one embodiment of the present application, the step of inputting the feature extraction vector into a classification network model, the classification network model classifying the feature extraction vector, and obtaining the classification result comprises:

[0118] S205, parsing the feature extraction vector to obtain a second feature vector;

[0119] S206: Input the second feature vector into a classification network model, and the classification network model classifies the second feature vector to obtain a classification result, wherein the classification result includes: a start forward state, a deceleration stop state, and a no state result category.

[0120] Please refer to Figure 1-Figure 3In this embodiment, the feature extraction vector obtained from the feature extraction module of the MobileNet neural network model is input into the classification module.

[0121] The second feature vector of the middle layer is extracted from the multi-layer feature vector as the input of the classification. The second feature vector has both details and global information, which is suitable for state classification. The second feature vector is input into the classification network model. The classification network performs feature recognition and classification of the state through operations such as convolutional layers and fully connected layers.

[0122] The output of the classification network model is divided into three status categories:

[0123] Start forward state: prompts the system that the vehicle in front has started to move, and the driver needs to pay attention. Slow down and stop state: prompts the system that the vehicle in front has slowed down or stopped, and the driver needs to keep a safe distance. No state: prompts the system that the current monitoring is invalid or the state cannot be determined, and enters the next round of detection.

[0124] When training the classification network model, images and annotation labels correspond one to one. The annotation labels only focus on the state of the center area in front of the image (including whether there is a vehicle and the driving status of the vehicle). During inference, if the output result is no target, no subsequent processing will be performed. If the output is start forward or slow down and stop, a corresponding prompt will be given.

[0125] By analyzing the feature extraction vector and selecting the second feature vector, the processing of redundant data is avoided and the classification efficiency is improved.

[0126] The three classification states cover the main scenarios of start-stop detection of the preceding vehicle, and can meet the application requirements under complex road conditions (such as congestion, following a vehicle, etc.). The classification network helps to reduce false alarms caused by false detection and improve the reliability of the system through precise feature analysis and state classification. The classification results provide a basis for generating driving prompt signals, helping drivers to understand the status of the preceding vehicle more intuitively, so that they can respond in time and reduce the risk of accidents. Through the streamlined design of the classification network, only the middle-level feature vectors are classified, avoiding the high computing power requirements of processing the full feature vector, which is particularly suitable for embedded and resource-limited vehicle systems.

[0127] The design of the classification network model is independent of the feature extraction module, which facilitates further optimization or replacement and is more adaptable to new state categories that may be added in the future.

[0128] It should be noted that the multi-level feature extraction vector obtained from the feature extraction module is parsed and the "second feature vector" of the middle layer is selected. The second feature vector is a specific layer feature extracted from the feature extraction vector after three downsamplings. Selecting feature vectors with medium semantic expression capabilities not only retains rich image information but also reduces redundancy and noise. Improve the efficiency and accuracy of the classification network and ensure that the input data is most suitable for the classification task. The parsed second feature vector is passed as input to the classification network model. The classification network model is a dedicated neural network module designed to identify the start-stop status of the preceding vehicle.

[0129] Based on the input of specific feature vectors, the classification network model can focus on identifying information related to the state of the preceding vehicle. This reduces the computational burden of the classification network, improves the real-time performance of the model, and optimizes the accuracy of state recognition. The classification network model maps the input feature vector to a predefined classification space through the operation of a multi-layer neural network, and finally outputs a specific state category. The state of the preceding vehicle is accurately classified so that the system can make corresponding prompts or decisions based on the state. The state of the preceding vehicle is clarified through the classification results, providing a direct decision-making basis for the driving assistance system.

[0130] In at least one embodiment of the present application, the step of inputting the feature extraction vector into a classification network model, the classification network model classifying the feature extraction vector, and obtaining the classification result includes:

[0131] S207, determining whether the classification result is stateless;

[0132] S208: If the classification result is no state, perform image processing in the next cycle.

[0133] Please refer to Figure 1-Figure 3 In this embodiment, the classification network model outputs the classification results of the front vehicle status, including "start forward state", "deceleration stop state" and "no state". The system logic module analyzes the classification results to determine whether it is "no state":

[0134] If the classification result is "no state": the current image pair cannot accurately reflect the status of the vehicle in front. If the classification result is other states: the system continues to perform subsequent depth calculations or prompt signal generation. If the classification result is "no state", the system skips the subsequent depth calculation and signal generation steps of the current cycle and directly enters the image processing of the next cycle.

[0135] In the next cycle of image processing, the image pair is re-acquired, the feature extraction vector and classification results are updated, and the state of the preceding vehicle is tried again. When the classification network determines that the state of the preceding vehicle is "stateless", the system actively skips the subsequent calculations and processing to avoid unnecessary resource consumption. This optimization is particularly suitable for embedded systems or scenarios with limited computing power. By entering the next cycle of processing, the system can more flexibly adapt to dynamic environmental changes and improve detection accuracy.

[0136] Through "stateless" result judgment, false reminders or misleading drivers caused by uncertain results are avoided, improving the reliability of the driving assistance system and user experience.

[0137] Entering the image processing process of the next cycle ensures that the system can update the status detection results at a faster frequency, especially in dynamic scenes, maintaining the system's real-time perception of the environment.

[0138] It should be noted that the system inputs the feature extraction vector into the classification network model, and the classification network model classifies the feature extraction vector and outputs clear classification results, including three possibilities: "start forward state", "deceleration stop state" and "no state". Obtain the state of the vehicle in front to provide a basis for the next logical judgment. The system performs logical judgment on the classification results to confirm whether the state of the vehicle in front is "no state". "No state" usually means that the vehicle in front is not moving. After determining that the classification result is "no state", the system does not perform depth calculation and prompt signal generation, but directly jumps to the image processing process of the next cycle. By entering the next cycle of image processing, the system's adaptability to complex dynamic scenes is improved (for example, classification anomalies caused by changes in light and viewing angle). Avoid deep calculation of misclassifications and reduce redundant operations, while ensuring that the system has continuity and responsiveness in a dynamic environment.

[0139] It needs to be further explained that there is a situation where the system makes a logical judgment on the classification result to confirm whether the status of the preceding vehicle is "stateless". "Stateless" usually means that the classification model cannot effectively distinguish the target in the current scene. After determining that the classification result is "stateless", the system does not perform depth calculation and prompt signal generation, but directly jumps to the image processing process of the next cycle. By entering the image processing of the next cycle, the system's adaptability to complex dynamic scenes is improved (for example, classification anomalies caused by changes in light and viewing angle). Improve the robustness of the system, and actively enter the next cycle for re-analysis when uncertainty is high to ensure the accuracy and reliability of the first prompt signal.

[0140] In at least one embodiment of the present application, the method further includes:

[0141] S209: If the classification result is determined to be a start forward state or a deceleration stop state (ie, the classification result is not a no state), a first prompt signal is generated according to the state corresponding to the classification result, and the first prompt signal is displayed on a vehicle display.

[0142] Please refer to Figure 1-Figure 3 In this embodiment, the classification network model outputs the state of the vehicle in front ("start forward state" or "decelerate and stop state"). The system performs a logical judgment on the classification result. If the judgment result is "start forward state", the system generates a first prompt signal of "the vehicle in front starts" and suggests the driver to follow the vehicle. If the judgment result is "decelerate and stop state", the system generates a first prompt signal of "the vehicle in front stops" to remind the driver to brake or keep a safe distance. The generated first prompt signal is transmitted to the vehicle display, and the display presents the prompt signal to the driver in the form of text or icon through a preset interface. The driver adjusts his driving behavior according to the prompt signal on the vehicle display, such as starting to follow the vehicle, decelerating or stopping. The first prompt signal is automatically generated and displayed to inform the driver of the status of the vehicle in front in real time, which significantly improves the intelligent level of driving assistance. An accurate first prompt signal is generated through the classification result to remind the driver to adjust his driving behavior in a complex traffic environment in a timely manner to avoid potential collision risks.

[0143] The first prompt signal converts the complex status detection results of the preceding vehicle into simple and intuitive visual prompts, reducing the driver's burden of observing and analyzing the traffic environment. In a variety of scenarios such as traffic congestion, traffic light intersections, and parking and following vehicles, the system can provide effective auxiliary reminders to improve the driving experience. An accurate first prompt signal can help the driver make correct decisions when information is incomplete or attention is distracted, reducing the probability of misoperation. The generation of the first prompt signal is triggered by the judgment of the classification results, and the status detection and prompt signal generation modules are separated, which enhances the scalability and maintainability of the system. The first reminder signal not only improves safety, but also avoids the driver's delayed response when the preceding vehicle starts, thereby improving driving efficiency.

[0144] It should be noted that logical judgment is performed through the output results of the classification network model ("start forward state" or "deceleration and stop state"). These two states correspond to two significant dynamic behaviors of the front vehicle: Start forward state: the front vehicle turns from a stationary state to forward. Deceleration and stop state: the front vehicle is decelerating or has completely stopped. Clarify the behavioral state of the front vehicle and determine whether it is necessary to prompt the driver. The logical judgment of the classification result provides the system with clear trigger conditions to ensure that the generation of the prompt signal is based on reliable state recognition and avoid unnecessary false alarms. Based on the determined classification results, a first prompt signal corresponding to the state of the front vehicle is generated:

[0145] Start forward state: Generates the first prompt signal of "the front vehicle has started" to remind the driver to follow.

[0146] Deceleration and stop state: Generates the first prompt signal of "the vehicle ahead has stopped" to remind the driver to keep a distance or prepare to stop.

[0147] The detected status of the vehicle ahead is intuitively transmitted to the driver through signal generation, assisting him to make quick decisions in different traffic scenarios. The generated prompt signal is visualized through the vehicle display. The display content can be text, icons or other forms, depending on the system design. Ensure that the driver can obtain the prompt information of the status of the vehicle ahead at the first time, and guide him to operate through visual prompts. Intuitive and clear visual display can effectively improve the driver's reaction speed and reduce the safety hazards that may be caused by delayed information transmission.

[0148] In at least one embodiment of the present application, the method further includes:

[0149] S301, if the classification result is determined to be a start-up forward state or a deceleration stop state, inputting the feature extraction vector into a depth calculation model;

[0150] S302: A deep calculation model analyzes the feature extraction vector to obtain a first feature vector, a second feature vector, and a third feature vector;

[0151] S303: The depth calculation model performs convolution calculation and upsampling on the third feature vector to obtain a first depth feature;

[0152] S304, the depth calculation model performs channel superposition on the second feature vector and the first depth feature, and then performs convolution calculation and upsampling to obtain a second depth feature;

[0153] S305, the depth calculation model performs channel superposition on the first feature vector and the second depth feature, and then performs convolution calculation and upsampling to obtain a depth feature vector;

[0154] S306. The depth calculation model uses a 1x1 convolution kernel to transform the depth feature vector to obtain a depth information value.

[0155] Please refer to Figure 1-Figure 3 In this embodiment, the system determines whether to start the depth calculation model by judging whether the classification result is "start forward state" or "deceleration stop state".

[0156] The deep computation model parses the input feature extraction vector, separates the first, second, and third feature vectors, and provides multi-scale input for subsequent deep computation.

[0157] The deep computing model processes feature vectors layer by layer:

[0158] Third feature vector: Generate the first deep feature through convolution and upsampling.

[0159] Second feature vector: superimposed with the first depth feature to generate the second depth feature.

[0160] First feature vector: superimposed with the second deep feature to generate the final deep feature vector.

[0161] The final depth feature vector is transformed through a 1x1 convolution kernel to output the depth information value, that is, the distance between the vehicle and the vehicle in front.

[0162] The depth information value is used as core data and input into the safety distance judgment module to generate safety warning signals or assist driving decisions.

[0163] Through multi-layer feature fusion and convolution calculation, accurate depth information values ​​are generated, which improves the reliability of depth perception.

[0164] The staged feature processing (layer-by-layer convolution and upsampling) avoids the high cost of global calculation and is particularly suitable for embedded systems with limited resources. Through the analysis and fusion of multiple layers of features, this method can extract accurate depth information in complex dynamic environments and adapt to various driving scenarios such as congestion and following vehicles.

[0165] The process is started only when deep calculation is needed to avoid invalid calculations and improve the utilization efficiency of system resources. The depth information value is the key data for generating safety warning signals, ensuring that the driver can accurately understand the distance from the vehicle in front and reduce the risk of traffic accidents. The modular design of the deep calculation model is easy to upgrade, such as introducing more feature layers or optimizing convolution methods to adapt to more complex driving scenarios in the future. Accurate deep calculation and timely warning signal generation improve the performance of the driving assistance system and provide drivers with a more intuitive and safe driving experience.

[0166] It should be noted that when the classification result is "start forward state" or "deceleration stop state", the system triggers the deep calculation process. At this time, the feature extraction vector generated by the classification network is input into the deep calculation model.

[0167] The depth calculation process is started only when a more accurate depth estimation of the distance to the vehicle ahead is required to avoid unnecessary calculations when there is no state result.

[0168] Optimize resource utilization, reduce the frequency of deep computing, reduce system computing power requirements, and ensure that deep computing focuses on critical states.

[0169] The deep computing model parses the input feature extraction vector and separates three levels of feature vectors:

[0170] First eigenvector: low-level features, containing detailed information.

[0171] The second feature vector: mid-level features, containing mid-level semantics.

[0172] The third feature vector: high-level features with global semantics.

[0173] Extract multi-level information of feature vectors to provide a comprehensive feature basis for subsequent depth inference.

[0174] Hierarchical feature parsing enables the model to focus on local details while capturing global context, providing multi-scale support for deep computing.

[0175] The high-level semantic features (third feature vector) are convolved to extract more accurate depth-related information. The higher spatial resolution is restored by upsampling to generate the first depth feature.

[0176] The high-level feature vectors are converted into deep features to capture the global depth information of the scene.

[0177] By processing high-level features, preliminary deep features are generated, laying the foundation for the gradual fusion and optimization of subsequent deep features. The second feature vector (middle-level feature) is channel-superimposed with the first deep feature to fuse the local semantics of the middle-level feature and the global depth information of the first deep feature, and then the second deep feature is generated through convolution and upsampling. The depth features are gradually refined while retaining the semantic information of the middle-level features. The middle-level features and global depth information are fused to generate more refined and accurate deep features. The low-level features (first feature vector) and the second deep feature are fused, and the final deep feature vector is generated through convolution and upsampling.

[0178] Comprehensively optimize the depth features by integrating the information of multiple layers of features to ensure high-resolution depth information output. Preserve the comprehensive information of scene details and global depth to provide accurate input for the final calculation of depth information value. Use 1x1 convolution kernel to reduce the dimension or map the depth feature vector, and finally generate a scalar depth information value. Extract the key information of the depth feature vector and calculate the actual distance (depth information value) between the vehicle and the vehicle in front. Efficiently complete the extraction of depth information to provide core data for the subsequent generation of safety warning signals or the execution of driving assistance decisions.

[0179] In at least one embodiment of the present application, the step of generating a safety distance signal according to the depth information value and displaying the safety distance signal on a vehicle display to remind the driver includes:

[0180] S307, calculating the distance value L to the preceding vehicle according to the depth information value d:

[0181] L = αd, where α is a constant;

[0182] S308. If the distance value L ≤ the first preset safety distance value c1, generate a first distance reminder signal;

[0183] S309. If the first preset safety distance value c1 < the distance value L ≤ the second preset safety distance value c2, generate a second distance reminder signal.

[0184] Please refer to Figure 1-Figure 3 , in this embodiment, the depth information value d obtained from the depth calculation model is input into the linear conversion formula L = αd to calculate the actual distance L between the vehicle and the vehicle in front.

[0185] The system compares the actual distance L with the first preset safety distance and the second preset safety distance:

[0186] If L ≤ the first preset safety distance: The system generates a first distance reminder signal indicating a higher risk.

[0187] If the first preset safety distance < L ≤ the second preset safety distance: The system generates a second distance reminder signal to prompt the driver to pay attention to the vehicle distance. According to the judgment result of the distance value, generate the corresponding safety distance reminder signal and convert it into a reminder form suitable for the driver to understand (such as sound, icon or text). The safety distance signal is visualized through the vehicle display, and the display transmits the signal to the driver in an intuitive form (such as color change, icon display or text warning). The driver adjusts the driving behavior according to the displayed safety distance signal, such as decelerating, braking or maintaining the vehicle distance. The depth information value is quickly converted into the actual distance through a simple linear formula to meet the real-time requirement and ensure the immediate generation of the reminder signal.

[0188] By distinguishing between high-risk and medium-risk states, the system classifies different danger levels, which not only improves the pertinence of the reminder but also avoids the fatigue that may be caused by frequent high-priority reminders. The first distance reminder signal can immediately warn the driver in a high-risk state and reduce the collision risk caused by too close a vehicle distance; the second distance reminder signal provides early warning for the driver under medium-risk conditions. The visualized reminder signal is presented intuitively through the vehicle display, which is easy for the driver to understand and operate, enhancing the user-friendliness of the system. The preset value of the safety distance can be flexibly adjusted according to the actual application scenario (such as urban driving, highway driving), enabling the system to function in multiple scenarios. The system actively reminds the driver of the vehicle distance situation, reducing the driver's continuous attention burden on the surrounding environment and improving driving efficiency and comfort. The signal generation logic is independent of the depth calculation module, and the preset safety distance threshold and reminder signal mechanism can be flexibly adjusted or extended according to specific application requirements.

[0189] It should be noted that, in this embodiment, for example, the first preset safety distance is set to 1.5, the second preset safety distance is set to 3, and a first threshold prompt is made when L≤1.5, and a second threshold prompt is made when 1.5<L≤3.

[0190] It should be further explained that the depth information value d is the depth data between the vehicle and the vehicle in front obtained from the depth calculation model. The depth information value is converted into the actual physical distance L through the linear conversion formula L=αd. When the distance L is less than or equal to the first preset safety distance value, a higher priority reminder signal (first distance reminder signal) is generated to warn the driver that there is a high risk of collision. The driver is reminded in time to take braking measures when too close to the vehicle in front to ensure safety. The generation of high-priority signals improves the system's responsiveness to emergencies, provides the driver with clear braking prompts, and reduces the risk of accidents.

[0191] If the distance L is between the first preset safety distance and the second preset safety distance, a reminder signal with a lower priority (second distance reminder signal) is generated to remind the driver to pay attention to the current vehicle distance condition.

[0192] Provides continuous distance reminders to the driver, especially in medium-risk conditions, to remind the driver to maintain an appropriate distance. The second distance reminder signal provides a progressive warning mechanism to prevent the driver from becoming fatigued or disgusted by frequent high-priority reminders. The first and second distance reminder signals are visualized through the vehicle display to remind the driver of the current distance status and necessary driving operations.

[0193] The safety distance signal is conveyed to the driver in an intuitive way, making it easier for the driver to quickly understand and take corresponding measures. Visual display significantly improves the user experience of the driving assistance system, allowing the driver to quickly perceive the dynamic environmental changes around the vehicle.

[0194] In at least one embodiment of the present application, the specific steps of acquiring a driving image of the vehicle during driving by a monocular camera in front of the vehicle and constructing an image pair based on the time sequence of the driving image include:

[0195] S201, acquiring a current frame image and a fourth frame image before the current frame image in the driving image, and superimposing the current frame image and the fourth frame image on a channel to obtain a superimposed image pair;

[0196] S202, performing image scaling and image normalization processing on the image pair to obtain a processed image pair;

[0197] S203: input the processed image pair into the MobileNet neural network model.

[0198] Please refer to Figure 1-Figure 2 In this embodiment, the monocular camera continuously captures the driving image in front of the vehicle. The system extracts the current frame image and the fourth frame image before it to form image data at two time points.

[0199] The current frame and the fourth frame are superimposed on the channel to form an image pair. Through this operation, the timing information is saved in a unified input data structure for subsequent processing.

[0200] The superimposed image pairs are scaled and normalized. Scaling ensures that the data fits the input size of MobileNet, and normalization keeps the data distribution consistent with the distribution during model training.

[0201] The preprocessed image pairs are input into the MobileNet neural network model. The model extracts multi-level feature vectors through its feature extraction module, providing basic data for subsequent classification or depth inference.

[0202] By selecting two frames of images with a large time interval and superimposing them, the ability to capture moving targets and scene changes is enhanced, providing a reliable basis for subsequent depth estimation and state recognition.

[0203] By selecting the fourth frame to be superimposed with the current frame instead of consecutive frames, redundant calculations caused by small changes between frames are reduced, while the system's sensitivity to larger dynamic scene changes is improved.

[0204] Image scaling and normalization operations make the data meet the input requirements of the model, ensure the stability of feature extraction, and avoid the degradation of model performance due to inconsistent data distribution.

[0205] The introduction of MobileNet ensures the system's operating efficiency in an embedded environment with limited computing power, while meeting the real-time requirements of the preceding vehicle start-stop detection.

[0206] Based on the acquisition of images by a monocular camera, no additional hardware equipment is required, which reduces the implementation cost and ensures that the system can be widely used in traditional and new vehicles.

[0207] The image pair after channel superposition not only retains the information independence of the two frames of images, but also enhances the adaptability of the model in different scenarios through standardization processing.

[0208] By integrating multi-frame information through channel superposition, it can be easily expanded to scenarios using more time frames to support more complex task requirements in the future.

[0209] It should be noted that two frames of images are extracted in continuous time from the monocular camera in front of the vehicle: the current frame image (reflecting the scene in front of the vehicle in real time) and the fourth frame image before (representing the scene a fixed time interval ago).

[0210] Using two frames of images with a large time interval, dynamic information in the scene is captured to provide a basis for inferring depth information. The frame difference in time can enhance the perception of moving targets or environmental changes, making up for the lack of dynamic information in a single frame. The two frames of images are superimposed in the channel dimension to form an image pair containing time sequence information. After superposition, each pixel retains the data of the two frames of images for subsequent feature extraction.

[0211] Through channel superposition, the information of two frames of images is combined to simplify the subsequent processing process and provide a unified input format for feature extraction.

[0212] It maintains the independence between the two frames of images to avoid confusion; at the same time, it provides an effective encoding form of timing information, which helps the subsequent model to extract dynamic features.

[0213] Image scaling adjusts the image to the input size required by the MobileNet neural network model, and image normalization standardizes the pixel values ​​to the range adapted by the model (usually [0, 1] or [-1, 1]).

[0214] In this embodiment, the monocular camera captures RGB three-channel images, and every three frames form an image pair (the current frame imgt0 and the past frame imgt-3 form a pair, and so on), and the image pairs are superimposed on the channel to form a new img. Since the input size of the feature extraction part is 288×448 (height×width), the image captured by the camera needs to be scaled to 288×448, and then normalized to -1 to 1. The normalization calculation method is: img=img / 127.5-1.

[0215] Through standardization, the computational efficiency of image pairs is improved, and the input data is ensured to be consistent with the distribution of model pre-training data. The impact of image size or pixel value differences on model reasoning accuracy is eliminated, and the adaptability and robustness of the neural network are enhanced. The processed image pairs are used as input to the MobileNet model. The MobileNet model is mainly characterized by lightweight and efficient feature extraction capabilities. The features in the image pairs are extracted through deep neural networks to provide high-dimensional expressions for subsequent classification and deep inference.

[0216] The MobileNet model provides reliable feature support for tasks such as object detection and state recognition in scenes through its hierarchical feature extraction capabilities, while also having the advantage of low computational cost.

[0217] A monocular camera-based front vehicle start / stop detection system 100 is applied to the monocular camera-based front vehicle start / stop detection method as described above, and the system includes:

[0218] An image acquisition module 110 is used to acquire an image from a monocular camera in front of the vehicle during driving;

[0219] MobileNet neural network model 120, performs feature extraction on the image pair to obtain a feature extraction vector;

[0220] The classification network model 130 classifies the input feature extraction vector to obtain a classification result;

[0221] The depth calculation model 140 calculates the feature extraction vector according to the classification result to obtain a calculated depth information value;

[0222] A safety distance generating module 150 generates a safety distance signal according to the depth information value;

[0223] The monocular camera-based front vehicle start / stop detection system 100 performs the following steps:

[0224] Acquire driving images of the vehicle during driving by using a monocular camera in front of the vehicle, and construct image pairs for the driving images according to a time sequence;

[0225] Input the image pair into the MobileNet neural network model 120, and the MobileNet neural network model 120 inputs the image pair into a feature extraction module to obtain a feature extraction vector;

[0226] Input the feature extraction vector into the classification network model 130, and the classification network model 130 classifies the feature extraction vector to obtain a classification result;

[0227] Input the classification result and the feature extraction vector into the depth calculation model 140 to calculate the depth information value;

[0228] A safety distance signal is generated according to the depth information value, and the safety distance signal is displayed on a vehicle display to remind the driver.

[0229] Please refer to Figure 4In this embodiment, the system obtains the real-time driving image in front of the vehicle collected by the monocular camera on the vehicle through the image acquisition module 100. The system generates an image pair through the time series construction technology to ensure that the input data has the potential for depth information. The system inputs the image pair into the MobileNet neural network model 120, and extracts feature vectors of different levels after three downsamplings. These features retain the key information of the motion state and scene of the front vehicle. The system inputs the extracted feature vector into the classification network model 130. The model classifies the state according to the image content to determine whether the front vehicle is started, stopped, or in a stateless state. The system inputs the classification result and the feature vector into the depth calculation model 140. The model generates a depth information value reflecting the distance of the front vehicle through layer-by-layer convolution calculation and upsampling. The system inputs the depth information value into the safety distance generation module 150, and generates safety distance signals of different levels in combination with the preset safety distance threshold, and reminds the driver through the vehicle display.

[0230] Using a monocular camera instead of a binocular camera or radar can significantly reduce hardware costs and simplify system installation and maintenance. The lightweight MobileNet neural network model and optimized deep computing model are used to significantly reduce computing power requirements, making the solution suitable for traditional models. The accuracy and reliability of the start-stop detection of the preceding vehicle are improved by linking the image pairs, classification network models, and deep computing models in a time-series manner. The detection solution for complex traffic scenarios such as congestion and following vehicles provides drivers with accurate safety distance information prompts. The safety distance signal is generated and displayed in a timely manner, which significantly improves the driver's reaction speed in complex road conditions and reduces the probability of accidents.

[0231] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0232] Acquire driving images of the vehicle during driving by using a monocular camera in front of the vehicle, and construct image pairs for the driving images according to a time sequence;

[0233] Inputting the image pair into a MobileNet neural network model, and the MobileNet neural network model inputting the image pair into a feature extraction module to obtain a feature extraction vector;

[0234] Inputting the feature extraction vector into a classification network model, the classification network model classifies the feature extraction vector to obtain a classification result;

[0235] Inputting the classification result and the feature extraction vector into a depth calculation model to calculate the depth information value;

[0236] A safety distance signal is generated according to the depth information value, and the safety distance signal is displayed on a vehicle display to remind the driver.

[0237] In this embodiment, the system obtains the real-time driving image in front of the vehicle collected by the monocular camera on the vehicle through the image acquisition module, and the system generates image pairs through time-series construction technology to ensure that the input data has the potential for depth information. The system inputs the image pairs into the MobileNet neural network model, and extracts feature vectors of different levels after three downsamplings. These features retain the key information of the motion state and scene of the front vehicle. The system inputs the extracted feature vectors into the classification network model. The model classifies the state according to the image content to determine whether the front vehicle is started, stopped, or in a stateless state. The system inputs the classification results and the feature vectors into the depth calculation model. The model generates a depth information value reflecting the distance of the front vehicle through layer-by-layer convolution calculation and upsampling. The system inputs the depth information value into the safety distance generation module, combines the preset safety distance threshold, generates safety distance signals of different levels, and reminds the driver through the vehicle display.

[0238] Using a monocular camera instead of a binocular camera or radar can significantly reduce hardware costs and simplify system installation and maintenance. The lightweight MobileNet neural network model and optimized deep computing model are used to significantly reduce computing power requirements, making the solution suitable for traditional models. The accuracy and reliability of the start-stop detection of the preceding vehicle are improved by linking the image pairs, classification network models, and deep computing models in a time-series manner. The detection solution for complex traffic scenarios such as congestion and following vehicles provides drivers with accurate safety distance information prompts. The safety distance signal is generated and displayed in a timely manner, which significantly improves the driver's reaction speed in complex road conditions and reduces the probability of accidents.

[0239] Figure 5 FIG. 1 shows an internal structure diagram of a computer device in an embodiment. The computer device may be a terminal or a server. Figure 5 As shown, the computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement a start-stop detection method for a preceding vehicle based on a monocular camera.

[0240] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0241] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for detecting the start and stop of a preceding vehicle based on a monocular camera, characterized in that: The method comprises: Acquire driving images of the vehicle during driving by using a monocular camera in front of the vehicle, and construct image pairs for the driving images according to a time sequence; Inputting the image pair into a MobileNet neural network model, and the MobileNet neural network model inputting the image pair into a feature extraction module to obtain a feature extraction vector; Inputting the feature extraction vector into a classification network model, the classification network model classifies the feature extraction vector to obtain a classification result; Inputting the classification result and the feature extraction vector into a depth calculation model to calculate the depth information value; A safety distance signal is generated according to the depth information value, and the safety distance signal is displayed on a vehicle display to remind the driver.

2. The method for detecting the start and stop of the preceding vehicle based on a monocular camera according to claim 1, characterized in that: The specific steps of inputting the image pair into the MobileNet neural network model, and the MobileNet neural network model inputting the image pair into the feature extraction module to obtain the feature extraction vector include: The feature extraction module downsamples the image pair three times to obtain feature extraction vectors after the three downsamplings, wherein the feature extraction vectors include: a first feature vector of the first downsampling, a second feature vector of the second downsampling, and a third feature vector of the third downsampling.

3. The method for detecting the start and stop of a preceding vehicle based on a monocular camera according to claim 1, characterized in that: The specific steps of inputting the feature extraction vector into the classification network model, and the classification network model classifying the feature extraction vector to obtain the classification result include: Analyzing the feature extraction vector to obtain a second feature vector; The second feature vector is input into a classification network model, and the classification network model classifies the second feature vector to obtain a classification result, wherein the classification result includes: a start forward state, a deceleration stop state, and a no state result category.

4. The method for detecting the start and stop of the preceding vehicle based on a monocular camera according to claim 3 is characterized in that: The step of inputting the feature extraction vector into the classification network model, and the classification network model classifying the feature extraction vector, and obtaining the classification result includes: It is determined whether the classification result is stateless. If the classification result is stateless, the image processing of the next cycle is performed.

5. The method for detecting the start and stop of the preceding vehicle based on a monocular camera according to claim 4 is characterized in that: The method further comprises: If the classification result is determined to be a start-up forward state or a deceleration stop state, a first prompt signal is generated according to the state corresponding to the classification result, and the first prompt signal is displayed on a vehicle display.

6. The method for detecting the start and stop of the preceding vehicle based on a monocular camera according to claim 4 is characterized in that: The method further comprises: If the classification result is determined to be a start-up forward state or a deceleration stop state, the feature extraction vector is input into a depth calculation model; The deep computing model analyzes the feature extraction vector to obtain a first feature vector, a second feature vector, and a third feature vector; The depth calculation model performs convolution calculation and upsampling on the third feature vector to obtain a first depth feature; The depth calculation model performs channel superposition on the second feature vector and the first depth feature and then performs convolution calculation and upsampling to obtain the second depth feature; The depth calculation model performs channel superposition on the first feature vector and the second depth feature and then performs convolution calculation and upsampling to obtain a depth feature vector; The depth calculation model uses a 1x1 convolution kernel to transform the depth feature vector to obtain a depth information value.

7. The method for detecting the start and stop of a preceding vehicle based on a monocular camera according to claim 1, characterized in that: The step of generating a safety distance signal according to the depth information value and displaying the safety distance signal on a vehicle display to remind the driver includes: The distance value L to the preceding vehicle is calculated based on the depth information value d: L = αd, where α is a constant; If the distance value L is less than or equal to the first preset safety distance value, a first distance reminder signal is generated; If the first preset safety distance value<distance value≤second preset safety distance value, a second distance reminder signal is generated.

8. The method for detecting the start and stop of a preceding vehicle based on a monocular camera according to claim 1, characterized in that: The specific steps of obtaining a driving image of the vehicle during driving by using a monocular camera in front of the vehicle and constructing an image pair based on the time sequence of the driving image include: Acquire a current frame image and a fourth frame image before the current frame image in the driving image, and superimpose the current frame image and the fourth frame image on a channel to obtain a superimposed image pair; Performing image scaling and image normalization processing on the image pair to obtain a processed image pair; The processed image pairs are input into the MobileNet neural network model.

9. A front vehicle start-stop detection system based on a monocular camera, applied to the front vehicle start-stop detection method based on a monocular camera as described in claims 1-8, characterized in that: The system comprises: An image acquisition module, used to acquire images from a monocular camera in front of the vehicle during driving; MobileNet neural network model, performs feature extraction on image pairs and obtains feature extraction vectors; The classification network model classifies the input feature extraction vector to obtain the classification result; The depth calculation model calculates the feature extraction vector according to the classification result to obtain the calculated depth information value; The safety distance generation module generates a safety distance signal according to the depth information value.

10. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 8.

Images