CN-121982647-A - Electric vehicle forbidden parking area supervision method and system

CN121982647ACN 121982647 ACN121982647 ACN 121982647ACN-121982647-A

Abstract

The invention discloses a supervision method and a supervision system for a forbidden parking area of an electric vehicle, and relates to the technical field of computer vision and smart city management. The system consists of four modules, namely image acquisition, improved target detection, violation judgment and communication reporting. The image acquisition module acquires a frame of static picture from fixed point monitoring every 5 seconds, the improved target detection module deploys an optimized YOLOv model, improves image quality under complex illumination through image preprocessing, adopts a GhostNet lightweight backbone network to reduce parameters and calculated amount, embeds a CBAM attention mechanism in a feature fusion network and combines a channel pruning strategy, accurately focuses a small-scale and low-contrast forbidden mark, and keeps high-efficiency reasoning, the violation determination module calculates target frame space overlapping degree to determine violation, and the communication reporting module uploads a result to the management platform. The system solves the problems of low model detection precision, slow response and high resource demand in the edge deployment scene, and improves the intelligent level and efficiency of urban forbidden management.

Inventors

WANG YAN
LIU CHEN
ZHANG ZONGTANG
LIU ANPENG
ZHANG JINGJING

Assignees

安徽理工大学

Dates

Publication Date: 20260505
Application Date: 20260130

Claims (9)

1. The method and the system for supervising the forbidden parking area of the electric vehicle are characterized by comprising the following steps: Collecting a static scene picture of a forbidden stop area of the electric vehicle; The static scene picture is subjected to depth separable convolution through a trunk network of an improved YOLOv model to extract multi-scale features, channel attention and space attention of a neck network of the improved YOLOv model are processed in parallel to fuse the multi-scale features and respond to features of small-scale stop marks, a detection head module of the improved YOLOv model outputs a detection frame list of an electric vehicle and a detection frame list of the stop marks, construction of the improved YOLOv model comprises adopting GhostNet as the trunk network of a native YOLOv model, and a CBAM attention module is embedded in a feature fusion network of the neck network of the native YOLOv model to obtain the improved neck network; based on the electric vehicle detection frame list and the forbidden stop identification detection frame list, obtaining the space overlapping degree between the electric vehicle detection frame and the forbidden stop identification detection frame; and packaging and reporting the electric vehicle violation judgment result to a remote management platform.
2. The method and system for supervising the forbidden parking area of the electric vehicle according to claim 1, wherein the image enhancement of the static scene picture comprises the following steps: and (3) carrying out automatic exposure AE and automatic white balance AWB adjustment on the static scene picture by calling a sensor of the camera so as to cope with complex illumination and environmental conditions such as dusk, night, rain and fog.
3. The method and system for supervision of an electric vehicle parking space of claim 1, wherein the modified YOLOv model further comprises: Sequentially performing contrast stretching and pixel normalization on the input picture; The contrast stretching, which linearly maps the pixel intensity value of the input picture from the original range [ a, b ] to the full range [0, 255] based on the image gray level histogram; the pixel normalization is carried out, and the pixel value of the image after the contrast stretching is normalized to a [0, 1] interval; The linear mapping, the functional expression is: ; The normalization is carried out according to the formula: ; wherein P represents a pixel value of the input picture; Representing the linearly mapped pixel values; Representing the normalized pixel values.
4. The method and system for supervising a forbidden parking area of an electric vehicle according to claim 1, wherein the depth separable convolution is performed on the static scene picture through a backbone network of a modified YOLOv model to extract multi-scale features, specifically comprising: based on the static scene picture, carrying out channel compression through standard 1x1 convolution to generate m pieces of characteristic feature images; Based on the separable convolution of the depth of 5x5, performing s-1 linear transformation on each intrinsic feature map respectively to generate s-1 phantom feature maps corresponding to the intrinsic feature maps; splicing the intrinsic feature map and all 'phantom' feature maps corresponding to the intrinsic feature map in the channel dimension to form a final output feature map; the eigenvector graph is expressed as: ; Generating s Zhang Huanying feature map for the ith feature map, which is expressed as: ; Wherein, the Y' _i represents each eigenvector graph, i=1, 2, & m, s represents the total eigenvector graph number corresponding to each eigenvector graph after phantom operation; the splicing in the channel dimension is expressed as: ; Wherein X represents the input enhanced image, y_ (i, j) represents the j-th ' phantom ' feature map generated by the j-th depth separable convolution operation of the i-th feature map Y ' _ i, and Y represents the final output feature map set.
5. The method and system for supervising a forbidden parking area of an electric vehicle of claim 1, wherein the CBAM attention module comprises: channel attention, carrying out global average pooling and global maximum pooling on an input feature map F at the same time, then adding two output vectors through a shared multi-layer perceptron MLP, and generating channel weights through a Sigmoid activation function ; The channel weight Expressed as: ; Spatial attention, respectively carrying out average pooling and maximum pooling on the characteristic F' defined by the channel attention in the channel dimension, splicing the results, and generating spatial weight through a standard convolution layer and a Sigmoid function ; The spatial weight Expressed as: ; final output after channel attention and spatial attention Expressed as: 。
6. The method and system for monitoring and controlling a parking area of an electric vehicle of claim 1, further comprising compressing the modified neck network by a channel pruning technique; performing sparsification training on the model by applying an L1 regularization penalty on the scaling factor gamma of the batch normalized BN layer in the C3K2 modules of the neck network; The loss function becomes: ; After model sparsification training is completed, counting absolute values of scaling factors gamma of all BN layers; And determining a global threshold of pruning, and for each BN layer, judging the convolution kernel corresponding to the channel with the scaling factor gamma value lower than the threshold as redundancy and pruning.
7. The method and system for supervising a forbidden parking area of an electric vehicle according to claim 1, wherein the spatial overlapping degree between the electric vehicle detection frame and the forbidden parking identification detection frame is calculated by the following steps: Traversing each electric vehicle detection frame ; For the current electric vehicle detection frame, traversing all forbidden stop mark detection frames ; Calculating electric vehicle frame With each forbidden stop identification frame Is a spatial overlap of (1); the spatial overlapping degree is calculated according to the following formula: Overlap = 。
8. The method and system for supervising the forbidden parking area of the electric vehicle according to claim 1, wherein the electric vehicle violation determination result comprises electric vehicle violation time, violation position and violation picture evidence.
9. An electric vehicle parking-forbidden area supervisory system, comprising: The image acquisition module is used for acquiring static scene pictures of the forbidden and stopped areas of the electric vehicle; The improved target detection module is used for carrying out depth separable convolution on the static scene picture through a main network of an improved YOLOv model to extract multi-scale features, carrying out parallel processing on channel attention and space attention of a neck network of the improved YOLOv model to fuse the multi-scale features and respond to features of small-scale stop marks, and outputting an electric vehicle detection frame list and a stop mark detection frame list through a detection head module of the improved YOLOv model, wherein the construction of the improved YOLOv model comprises adopting GhostNet as the main network of a native YOLOv model, and embedding the CBAM attention module into a feature fusion network of the neck network of the native YOLOv model to obtain the improved neck network; The system comprises an electric vehicle detection frame list, a rule breaking judgment module and a rule breaking judgment module, wherein the electric vehicle detection frame list and the rule breaking judgment module are used for obtaining the space overlapping degree between the electric vehicle detection frame and the rule breaking identification detection frame based on the electric vehicle detection frame list and the rule breaking identification detection frame list; And the communication reporting module is used for packaging and reporting the electric vehicle violation judgment result to the remote management platform.

Description

Electric vehicle forbidden parking area supervision method and system Technical Field The application relates to the technical field of computer vision and smart city management, in particular to a method and a system for supervising a forbidden stop zone of an electric vehicle. Background With the rapid increase of the conservation quantity of electric vehicles, the problem of illegal parking of the electric vehicles has become a prominent hidden trouble of urban management and fire safety. Especially in legal forbidden areas such as fire control passageway, pavement, emergent export, the electric motor car is in disorder and is put in disorder not only destroys urban order, more directly blocks up emergent rescue passageway, probably delays the calamity and handles the opportunity, constitutes serious threat to public security. Currently, electric vehicle forbidden supervision mainly depends on two traditional modes of manual patrol and fixed-point video monitoring. The fixed point monitoring can realize all-weather coverage, but relies on manual real-time staring or post playback checking, so that the instant supervision requirement of 'discovery and disposal' is difficult to meet, and the overall supervision efficiency is low. In order to solve the above problems, computer vision technology is widely used in the field of vehicle violation detection, and related technical solutions have been disclosed in several existing patents, but when adapting to the specific scenario of "electric vehicle forbidden stop", there is still an obvious technical short board. In recent years, a large model detection scheme based on a Transformer architecture is applied to vehicle violation identification (an intelligent traffic violation detection method based on an improved Transformer of publication number CN116204351 a), while global feature understanding capability can be improved, the number of model parameters is generally more than 1 hundred million, and a great amount of calculation force is consumed for single frame reasoning. If the scheme is deployed in low-cost equipment (such as embedded cameras and small edge boxes) of the edge scenes of communities, buildings and the like, high-performance hardware needs to be additionally configured, so that the deployment cost is increased by 3-5 times, and large-scale economic application is difficult to realize. Disclosure of Invention Based on the above, it is necessary to provide a method and a system for monitoring the forbidden parking area of an electric vehicle. The technical scheme adopted in the specification is as follows: the specification provides a supervision method for a forbidden parking area of an electric vehicle, which comprises the following steps: Collecting a static scene picture of a forbidden stop area of the electric vehicle; The static scene picture is subjected to depth separable convolution through a trunk network of an improved YOLOv model to extract multi-scale features, channel attention and space attention of a neck network of the improved YOLOv model are processed in parallel to fuse the multi-scale features and respond to features of small-scale stop marks, a detection head module of the improved YOLOv model outputs a detection frame list of an electric vehicle and a detection frame list of the stop marks, construction of the improved YOLOv model comprises adopting GhostNet as the trunk network of a native YOLOv model, and a CBAM attention module is embedded in a feature fusion network of the neck network of the native YOLOv model to obtain the improved neck network; based on the electric vehicle detection frame list and the forbidden stop identification detection frame list, obtaining the space overlapping degree between the electric vehicle detection frame and the forbidden stop identification detection frame; and packaging and reporting the electric vehicle violation judgment result to a remote management platform. Further, the image enhancement of the static scene picture specifically includes: and (3) carrying out automatic exposure AE and automatic white balance AWB adjustment on the static scene picture by calling a sensor of the camera so as to cope with complex illumination and environmental conditions such as dusk, night, rain and fog. Further, the improved YOLOv model further includes: Sequentially performing contrast stretching and pixel normalization on the input picture; The contrast stretching, which linearly maps the pixel intensity value of the input picture from the original range [ a, b ] to the full range [0, 255] based on the image gray level histogram; the pixel normalization is carried out, and the pixel value of the image after the contrast stretching is normalized to a [0, 1] interval; The linear mapping, the functional expression is: ; The normalization is carried out according to the formula: ; wherein P represents a pixel value of the input picture; Representing the linearly mapped pixel values; Representing