Parking space detection method and system based on information fusion

Abstract

The application discloses a kind of based on information fusion's parking stall detection method and system, detection method includes the position information of longitudinal and transverse distance of parking stall when seeking library collection;Collect and detect the image information of wheel hub, vehicle body and license plate of parking stall obstacle vehicle;Image information is detected using deep learning algorithm, the characteristic information of wheel hub, vehicle body and license plate of obstacle vehicle is determined;According to position information and characteristic information, the space position information of obstacle vehicle's vehicle body posture parameter and the parking stall to be parked is calculated;Parking stall space model is constructed and is planned out parking path, then signal is output to ESP chassis control system and EPS system, control vehicle tracks planned path and enters parking stall;The application improves prediction accuracy, and the generalization of model is good and robustness is stronger;While prediction speed is greatly accelerated, solves the problem that CenterNet original detector overflows in memory on commonly used embedded device, cannot train, enhances practicality.

Description

Technical Field

[0001] This invention belongs to the fields of machine learning and automotive electronics technology, and in particular relates to a parking space detection method and system based on information fusion. Background Technology

[0002] Automated parking systems are a current research hotspot in the automotive electronics field. The core technical challenge of automated parking systems lies in the detection of parking spaces and the determination of their type. The key to parking space detection is the perception of the surrounding environment, commonly achieved using sensors such as ultrasonic radar and cameras. Compared to lidar and cameras, ultrasonic radar is less expensive, less affected by weather, and readily obtains depth information. Cameras offer advantages such as lower cost, mature technology, and richer information acquisition. However, they are significantly affected by lighting and environmental conditions, making all-weather operation difficult. Therefore, current parking systems often employ a fusion of ultrasonic radar and cameras to form a perception module, such as using multiple ultrasonic sensors and a 360° surround-view video system to jointly extract parking space information.

[0003] Vehicle-mounted ultrasonic radar emits ultrasonic waves through an ultrasonic transmitter and calculates distance by receiving the time difference of the reflected waves. Commonly used probes operate at frequencies of 40kHz, 48kHz, and 58kHz. A 40kHz probe is frequently used to strike a balance between vertical distance and detection angle. Ultrasonic radar constructs a spatial model of a parking space by collecting depth information during the parking space search process. However, using ultrasonic radar alone can lead to difficulties in identifying obstacle types and recognition in open parking scenarios. Therefore, it is often combined with camera information for detection.

[0004] Because cameras lack depth information, current applications of automotive cameras in parking space detection modules primarily focus on image processing, extracting features from the images to obtain information about the surrounding environment. This extraction of environmental features involves various algorithms targeting different characteristics. One approach uses image recognition technology to directly identify obstacles and vehicles, while another uses camera image information to design specific algorithms for identifying parking lines. However, cameras are significantly affected by environmental factors; different angles, lighting conditions, and visibility all greatly impact recognition performance. Furthermore, the real-time requirements for data acquisition are very high in the field of autonomous driving, making algorithm speed a crucial indicator. Therefore, enhancing the generalization ability of the algorithm, ensuring effectiveness in different environments, and improving the algorithm's operating efficiency are key factors in the design of visual recognition algorithms.

[0005] The main approaches to fusing parking space visual detection algorithms with ultrasonic radar in existing technologies include:

[0006] 1. First, the images acquired from the panoramic vehicle imaging system undergo a series of preprocessing steps, including grayscale conversion, top-hat transformation, and histogram equalization, to transform overly bright or dark images into images with appropriate brightness, thereby enhancing the algorithm's anti-interference ability. Then, image filtering and sharpening are performed to extract edges and obtain typical features of parking lines and information within the parking spaces. Next, Radon transform is performed to obtain the point set information of the parking lines in Radon space. Finally, machine learning algorithms such as cluster analysis are used to achieve pixel-level localization of the parking lines. This method primarily relies on visual recognition algorithms, obtaining spatial information of parking spaces through pixel-level localization of parking lines. However, this method is not stable under poor environmental conditions such as heavy fog, and it cannot properly detect parking spaces when there are no parking lines or the parking lines are not clearly visible.

[0007] 2. Deep learning methods are used for vehicle recognition and obstacle detection. Currently, deep learning-based object detection algorithms are mainly divided into first-order and second-order detectors. Commonly used second-order detectors, such as Faster-RCNN, achieve high-accuracy object detection performance, and their advantages are very obvious compared to first-order detectors in multi-scale and small object problems. However, the improved accuracy brings the problem of slower speed and the real-time performance is not yet guaranteed in practical problems. First-order detectors, such as the classic algorithm YOLOv3, can achieve a forward speed of 3 to 4 times that of other detection algorithms when the accuracy is the same. However, in parking scenarios, such as dense scenes like parking lots, there will be problems of missed detections and false detections, and the safety of second-order detectors is lower than that of second-order detectors in the field of autonomous driving. Summary of the Invention

[0008] The purpose of this invention is to provide a parking space detection method and system based on information fusion to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a parking space detection method based on information fusion, comprising:

[0010] During the vehicle's search for parking spaces, the longitudinal and lateral distances are collected in real time.

[0011] Collect and detect image information of the wheel hubs, body, and license plates of vehicles obstructing parking spaces;

[0012] Import image information into a dataset, which includes a validation set and a test set;

[0013] The deep learning algorithm of CenterNet network is used to detect objects in the image information and determine the feature information of the wheel hub, body and license plate of the obstructing vehicle. The feature information includes the order and number of wheel hub, body and license plate of the obstructing vehicle in the parking space during the search process.

[0014] The vehicle body posture parameters of the obstacle vehicle and the spatial location information of the parking space are calculated based on the location information and feature information.

[0015] Construct a parking space model in the reference frame;

[0016] The parking path is planned based on the spatial model, and then the bus outputs the signal to the ESP chassis control system and EPS system to control the vehicle to follow the planned path and park in the parking space.

[0017] Furthermore, when the first and second transition points of the obstacle vehicle are detected, if the image information of the two wheel hubs is detected, the body attitude angle of the obstacle vehicle is determined to be an acute angle α2; if the license plate of the obstacle vehicle is detected between the two transition points but the wheel hub is not detected, the body attitude angle of the obstacle vehicle is determined to be an obtuse angle α1; then the minimum distance between the two obstacle vehicles is calculated to construct a parking space model.

[0018] Furthermore, during the vehicle's search for parking spaces, the longitudinal and lateral distance information of the parking spaces is collected in real time, and the parking space spacing parameter data L_min and D_min are fused together to formulate a rule base for determining the parking space type and establish a parking space space model. At the same time, the membership functions of the input parameters α1, α2, L_min and D_min of the fuzzy inference model are determined. After importing the membership functions into the fuzzy inference model, the parking space type can be determined based on the input. Here, L_min is the minimum lateral distance between vehicles with obstacles, and D_min is the minimum distance between vehicles with obstacles.

[0019] When D_min≥M+a, the parking space type is angled parking space; when D_min<M+a, the parking space type is non-parking space.

[0020] When L_min≥N+b, the parking space type is parallel parking space; when L_min<N+b, the parking space type is non-parking space.

[0021] When L_min≥M+a, the parking space type is perpendicular parking space; when L_min<M+a, the parking space type is non-parking space.

[0022] Where M is the width of the vehicle, N is the length of the vehicle, a is the width margin after the vehicle enters the parking space, and b is the length margin after the vehicle enters the parking space.

[0023] On the other hand, a parking space detection system based on information fusion is proposed, including:

[0024] The perception module collects positional information about the longitudinal and lateral distances to the obstacle vehicle, as well as image information about the obstacle vehicle's wheel hubs, vehicle body, and license plate.

[0025] The detection module performs target detection on the images of the wheel hubs, body, and license plates of the vehicles with obstacles and determines the feature information of the wheel hubs, body, and license plates.

[0026] The storage module is used to store the dataset of image information collected by the perception module. The dataset includes images of obstacle vehicles and images that have undergone data augmentation.

[0027] The calculation module determines the vehicle body posture parameters of the obstacle vehicle and the spatial location information of the parking space by calculating the target detection results, and then constructs a parking space spatial model under a defined reference frame.

[0028] Furthermore, the sensing module includes an ultrasonic radar, a wheeled odometer, and an onboard camera; the ultrasonic radar is used to collect longitudinal distance position information in real time; the wheeled odometer is used to collect lateral distance position information in real time; and the onboard camera is used to collect image information of the wheel hubs, body, and license plate of the obstacle vehicle.

[0029] Furthermore, the dataset contains 16,185 images, categorized into three types: the body, wheel rim, and license plate of the obstacle vehicle.

[0030] Furthermore, the detection module includes a CenterNet network, which is used for target detection.

[0031] Furthermore, the CenterNet network includes a feature extraction network, which is a 50-layer ResNet network.

[0032] Furthermore, the ResNet network architecture consists of four large convolutional groups, followed by a global average pooling to reduce the feature map size to 1×1, then full connectivity of the required dimensions, and finally a softmax function to output the classification score. For the recognition of wheel hubs, body and license plates of vehicles with obstacles in parking scenarios, the neural network is trained using the public dataset CarsDatasetss.

[0033] Furthermore, each convolutional group includes a Basicblock module and a Bottleneck module, which respectively indicate whether the dimensions of the input and output of the residual unit are the same, modify the output dimension of the fully connected layer so that it outputs the scores of the three categories after passing through the softmax layer.

[0034] Compared with existing technologies, the advantages of this invention are as follows: It utilizes a dataset of 16,185 images, with only three classification categories: vehicles, wheel rims, and license plates. Therefore, the prediction accuracy can reach nearly 99%. Furthermore, the dataset includes vehicle images under various environmental conditions and data-augmented images, ensuring good generalization and robustness of the model. Simultaneously, the use of a lightweight network, ResNet-50, combined with CenterNet significantly accelerates the prediction speed, reducing the prediction time for a single image to approximately 50ms. This solves the problem of memory overflow and training failure of the native CenterNet detector on common embedded devices, enhancing its practicality. Moreover, a fuzzy inference model is built; by adjusting the membership functions of various feature parameters, the type of parking space can be output from the input sensor data. Attached Figure Description

[0035] Figure 1 This is a flowchart of a parking space detection method based on information fusion in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the connection of a parking space detection system based on information fusion in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of database searching in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the obstacle vehicle's position in an embodiment of the present invention;

[0039] Figure 5 This is a diagram of the feature extraction network structure in an embodiment of the present invention;

[0040] Figure 6 This is a prediction accuracy graph from an embodiment of the present invention;

[0041] Figure 7 This is a fuzzy inference rule diagram in an embodiment of the present invention;

[0042] Figure 8 This is a membership function graph of each feature parameter in the embodiments of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0045] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0046] Please refer to the accompanying drawings in the specification. This invention provides a technical solution: such as... Figure 1 As shown, a parking space detection method based on information fusion includes the following steps:

[0047] S100: Real-time collection of longitudinal and lateral distance information of parking spaces during vehicle search process;

[0048] S200: Collect and detect image information of the wheel hubs, body, and license plates of vehicles obstructing parking spaces;

[0049] S300. Import the image information into the dataset, wherein the dataset includes a validation set and a test set;

[0050] S400: The deep learning algorithm of CenterNet network is used to perform target detection on image information and determine the feature information of the wheel hub, body and license plate of the obstructing vehicle. The feature information includes the order and number of wheel hub, body and license plate of the obstructing vehicle in the parking space during the search process.

[0051] S500: Calculates the vehicle body posture parameters of the obstacle vehicle and the spatial location information of the parking space based on the location information and feature information.

[0052] S600. Construct a parking space model in the reference frame;

[0053] The S700 plans a parking path based on the spatial model, and then outputs signals to the ESP chassis control system and EPS system via the bus to control the vehicle to follow the planned path and park in the parking space.

[0054] In the above embodiments, addressing the difficulty of achieving a balance between prediction speed and accuracy in classic deep learning algorithms, this invention employs the CenterNet network proposed in a 2019 CVPR paper. CenterNet is an anchor-free object detection network that abandons the anchor-point mechanism for generating candidate boxes, directly predicting the location of the object's center point, thus significantly improving object detection speed while also demonstrating good performance in detection accuracy. Figure 5 As shown, the ratio of validation set to test set was selected as 1:9 in the dataset, and the forward prediction accuracy was above 98%, with a convergence round count of around 0.2, demonstrating good accuracy and speed.

[0055] like Figure 4 As shown, optionally, when the first transition point A1 (A2) and the second transition point D1 (B2) of the obstacle vehicle are detected, if the image information of the two wheel hubs is detected, the body posture angle α of the obstacle vehicle is determined to be an acute angle α2; if the license plate of the obstacle vehicle is detected between the two transition points but the wheel hub is not detected, the body posture angle α of the obstacle vehicle is determined to be an obtuse angle α1; then the minimum distance between the two obstacle vehicles is calculated, and a parking space model is constructed.

[0056] In the above embodiments, the ultrasonic radar obtains parking space depth information based on the jump points of the detection distance, the odometer obtains width information, and the camera assists in judging the vehicle's attitude based on the number of wheel hubs and whether a license plate is detected. When the ultrasonic radar detects the vehicle's first jump point A1 (A2) and second jump point D1 (B2), if the camera identifies two wheel hubs, the vehicle's attitude angle α can be determined to be an acute angle. If a license plate is detected between the two jump points but no wheel hubs are detected, the attitude angle α can be determined to be an obtuse angle. The minimum distance between the two spaces can then be calculated using a formula to construct a parking space model. Figure 2 As shown, the parking controller plans the parking path based on the spatial model, and then the bus outputs the signal to the ESP chassis control system and EPS system to control the vehicle to follow the planned path and park in the parking space.

[0057] Optionally, such as Figure 3 , 4 As shown in Figure 8, during the vehicle's search for parking spaces, the longitudinal and lateral distance information of the parking spaces is collected in real time. The parking space spacing parameter data L_min and D_min are then fused together to obtain the parking space type. A rule base for determining the parking space type is formulated, and a parking space model is established. At the same time, the membership functions of the input parameters α1, α2, L_min and D_min of the fuzzy inference model are determined. After the membership functions are imported into the fuzzy inference model, the parking space type can be determined based on the input. Here, L_min is the minimum lateral distance between vehicles with obstacles, and D_min is the minimum distance between vehicles with obstacles.

[0058] like Figure 7 As shown, when D_min≥M+a, the parking space type is angled parking space; when D_min<M+a, the parking space type is non-parking space.

[0059] When L_min≥N+b, the parking space type is parallel parking space; when L_min<N+b, the parking space type is non-parking space.

[0060] When L_min≥M+a, the parking space type is perpendicular parking space; when L_min<M+a, the parking space type is non-parking space.

[0061] Where M is the width of the vehicle, N is the length of the vehicle, a is the width margin after the vehicle enters the parking space, and b is the length margin after the vehicle enters the parking space.

[0062] In the above embodiment, a is the distance between the vehicle and the vehicles on both sides after entering the parking space, which is a margin for activities such as opening and closing doors and preventing collisions with the vehicles on both sides; b is the distance between the vehicle and the vehicles in front and behind after entering the parking space, which is a margin for parking, driving out and preventing collisions with the vehicles in front and behind. Preferably, a = 0.6m and b = 0.8m.

[0063] On the other hand, such as Figure 2 As shown, a parking space detection system based on information fusion is proposed, including:

[0064] The perception module collects positional information about the longitudinal and lateral distances to the obstacle vehicle, as well as image information about the obstacle vehicle's wheel hubs, vehicle body, and license plate.

[0065] The detection module performs target detection on the images of the wheel hubs, body, and license plates of the vehicles with obstacles and determines the feature information of the wheel hubs, body, and license plates.

[0066] The storage module is used to store the dataset of image information collected by the perception module. The dataset includes images of obstacle vehicles and images that have undergone data augmentation.

[0067] The calculation module determines the vehicle body posture parameters of the obstacle vehicle and the spatial location information of the parking space by calculating the target detection results, and then constructs a parking space spatial model under a defined reference frame.

[0068] Optionally, the sensing module includes an ultrasonic radar, a wheeled odometer, and an onboard camera; the ultrasonic radar is used to collect longitudinal distance position information in real time; the wheeled odometer is used to collect lateral distance position information in real time; and the onboard camera is used to collect image information of the wheel hubs, body, and license plate of the obstacle vehicle.

[0069] In the above embodiments, the hardware of the present invention mainly consists of four ultrasonic radars and one forward-looking camera. During the parking space search process, the ultrasonic radars obtain parking space depth information based on the jump points of the detection distance, the odometer obtains width information, and the camera assists in judging the vehicle posture based on the number of wheel hubs and whether a license plate is detected.

[0070] Optionally, the dataset contains 16,185 images and is categorized into three types: the body, wheel rim, and license plate of the obstacle vehicle.

[0071] In the above embodiment, a dataset with 16,185 images and only 3 classification categories is used, so the prediction accuracy can reach close to 99%. In addition, the dataset contains vehicle images under various environmental conditions and includes data-enhanced images, so the generalization ability of the model can be well guaranteed and the robustness is strong.

[0072] Optionally, the detection module includes a CenterNet network for target detection.

[0073] Optionally, the CenterNet network includes a feature extraction network, which is a 50-layer ResNet network.

[0074] In the above embodiments, the original CenterNet detector uses the Hourglass-104 feature extraction network as its backbone, achieving prediction accuracy no less than that of a second-order detector. However, in automatic parking systems, Hourglass-104 suffers from a large number of parameters and high network complexity, resulting in slow detection speed. To address the problems of CenterNet in vehicle applications, an improved method is proposed: replacing the backbone feature extraction network with a 50-layer ResNet network. This lightweight backbone effectively solves the slow detection speed problem, significantly accelerating prediction speed, compressing the prediction time for a single image to approximately 50ms. This also resolves the memory overflow and training failure issues of the native CenterNet detector on common embedded devices, enhancing its practicality and improving the feasibility of CenterNet in parking applications.

[0075] Optionally, the ResNet network architecture consists of four large convolutional groups, followed by a global average pooling to reduce the feature map size to 1×1, then full connectivity of the required dimensions, and finally a softmax function to output the classification score. For the recognition of wheel hubs, body and license plates of vehicles with obstacles in parking scenarios, the neural network is trained using the public dataset CarsDatasetss.

[0076] Optionally, such as Figure 6 As shown, each convolutional group includes a Basicblock module and a Bottleneck module, which respectively indicate whether the dimensions of the input and output of the residual unit are the same, and modify the output dimension of the fully connected layer so that it outputs the scores of the three categories after passing through the softmax layer.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A parking space detection method based on information fusion, characterized in that, The application comprises the following steps: Collecting the position information of the longitudinal distance and the lateral distance of the parking space in the vehicle searching process; Collecting and detecting the image information of the hub, the body and the license plate of the obstacle vehicle in the parking space; Importing the image information into a data set, wherein the data set comprises a verification set and a test set; Using the deep learning algorithm of the CenterNet network to detect the image information and determine the feature information of the hub, the body and the license plate of the obstacle vehicle, wherein the feature information comprises the order and the number of the hub, the body and the license plate of the obstacle vehicle detected in the searching process; Calculating the body posture parameters of the obstacle vehicle and the spatial position information of the parking space to be parked according to the position information and the feature information; Constructing a parking space spatial model under the reference system; Planning a parking path according to the spatial model, and then outputting the signal to the ESP chassis control system and the EPS system through the bus to control the vehicle to track the planned path and park in the parking space; When the first and second jump points of the obstacle vehicle are detected, if the image information of two hubs is detected, it is determined that the body posture angle of the obstacle vehicle is an acute angle α2; if the license plate of the obstacle vehicle is detected between the two jump points but no hub is detected, it is determined that the body posture angle of the obstacle vehicle is an obtuse angle α1; and then the minimum distance between the two obstacle vehicles is calculated to construct the parking space spatial model.

2. The information fusion based parking space detection method according to claim 1, characterized in that, In the vehicle searching process, the longitudinal and lateral distance position information of the parking space is collected in real time, the distance parameter data L_min and D_min are obtained by fusion, the rule base for determining the type of parking space is formulated, the parking space spatial model is established, the membership functions of the input parameters α1, α2, L_min and D_min of the fuzzy reasoning model are determined, the membership functions are imported into the fuzzy push model, and then the type of parking space is determined according to the input, wherein L_min is the minimum lateral distance between the obstacle vehicles, and D_min is the minimum distance between the obstacle vehicles; When D_min is greater than or equal to M+a, the type of parking space is a diagonal parking space; when D_min is less than M+a, the type of parking space is a non-parking space; When L_min is greater than or equal to N+b, the type of parking space is a parallel parking space; when L_min is less than N+b, the type of parking space is a non-parking space; When L_min is greater than or equal to M+a, the type of parking space is a vertical parking space; when L_min is less than M+a, the type of parking space is a non-parking space; Wherein, M is the width of the vehicle, N is the length of the vehicle, a is the excess of the width direction when the vehicle enters the parking space, and b is the excess of the length direction when the vehicle enters the parking space.

3. A parking space detection system based on information fusion, characterized in that The application comprises the following steps: The perception module collects the position information of the longitudinal distance and the lateral distance of the obstacle vehicle, and the image information of the hub, the body and the license plate of the obstacle vehicle; The detection module detects the image information of the hub, the body and the license plate of the obstacle vehicle and determines the feature information of the hub, the body and the license plate; The storage module is used for storing the data set of the image information collected by the perception module, wherein the data set comprises the pictures of the obstacle vehicle and the pictures after data enhancement. The computing module determines the body posture parameter of the obstacle vehicle and the spatial position information of the parking space to be parked through the calculation of the target detection result, and further constructs a parking space spatial model under the reference system, wherein when the first and second jump points of the obstacle vehicle are detected, if the image information of the two hubs is detected, it is determined that the body posture angle of the obstacle vehicle is an acute angle alpha2; if the license plate of the obstacle vehicle is detected between the two jump points and the hub is not detected, it is determined that the body posture angle of the obstacle vehicle is an obtuse angle alpha1; and the minimum distance between the two obstacle vehicles is calculated to construct a parking space spatial model.

4. The information fusion based parking space detection system according to claim 3, characterized in that, The perception module includes an ultrasonic radar, a wheel odometer and a vehicle-mounted camera; the ultrasonic radar is used to collect real-time longitudinal distance position information; the wheel odometer is used to collect real-time transverse distance position information; and the vehicle-mounted camera is used to collect image information of the hub, body and license plate of the obstacle vehicle.

5. The information fusion based parking space detection system according to claim 3, characterized in that, The data set has 16185 pictures, and the classification categories are three, which are the body, hub and license plate of the obstacle vehicle.

6. The information fusion based parking space detection system according to claim 3, characterized in that, The detection module includes a CenterNet network, which is used for target detection.

7. The information fusion based parking space detection system according to claim 6, characterized in that, The CenterNet network includes a feature extraction network, which is a 50-layer ResNet network.

8. The information fusion based parking space detection system according to claim 7, characterized in that, The network architecture of the ResNet network is composed of four large convolution groups, and then the feature map size is changed to 1x1 through a global average pooling, and then the required dimension full connection is performed, and finally the classification score is output through the softmax function. For the identification of the hub, body and license plate of the obstacle vehicle in the parking scene, the Cars Datasetss public data set is used for neural network training.

9. The information fusion based parking space detection system according to claim 8, characterized in that, Each convolution group includes a Basicblock module and a Bottleneck module, which respectively represent whether the input and output dimensions of the residual unit are the same, modify the output dimension of the full connection layer, so that three categories of scores are output after the softmax layer.

Images