CN105760847A - Visual detection method for detecting whether motor cyclist wears helmet or not - Google Patents

Abstract

The present invention relates to a visual detection method for motorcycle drivers wearing helmets. The proposed method is divided into two stages: motorcycle vehicle segmentation and classification (used to judge whether the target concerned is a motorcycle) and whether the helmet is worn or not. detection. For motorcycle classification, the commonly used Haar features are used as descriptors and SVM classifiers; for helmet detection, we use circular hough transform (CHT) to extract the head region, and then use the histogram of orientation gradient (HOG) descriptor Extract image features, and classify objects with a multi-layer neural network (MLP) classifier; this method has high detection accuracy, good real-time performance, and has engineering practical value.

Description

A visual inspection method for motorcycle driver's helmet wearing

technical field

The invention belongs to the technical field of image processing, and relates to image segmentation and image information acquisition, in particular to a visual detection method for helmet wearing of motorcycle drivers, which is mainly used in road traffic safety monitoring and management.

Background technique

Traffic safety has become a major worldwide issue, and the impact of motorcycle safety on human life and property is self-evident. In many countries, with the increase in the number of motorcycles used, their driving speed has also increased accordingly. In addition to the increasingly busy traffic, motorcycle accidents caused by not wearing helmets have increased, causing a large number of casualties and property losses. The phenomenon has become a social problem that cannot be ignored. At present, the supervision of motorcycle drivers not wearing helmets, which only relies on the manual judgment of the traffic police, has become increasingly difficult to meet the growing number of motorcycles and the illegal behavior of not wearing helmets. Therefore, how to use extensive road monitoring and visual sensing to solve the problem of not wearing a helmet detection due to its large amount of information and low cost is of great significance.

Helmet detection technology refers to the process of using image sensing means to search and judge the driver in the image to decide whether to wear a helmet. At present, the technology and methods in this area are relatively lacking. Aiming at this, the present invention proposes a helmet detection computer vision system for motorcyclists traveling on public roads.

In order to describe the content of the present invention conveniently, some concepts need to be explained first.

Region of Interest: In the field of image processing, a region of interest refers to a local image region selected from an image, which is the focus of image analysis. Identify this area for further processing. Using regions of interest can often reduce processing time and increase accuracy.

Contents of the invention

The present invention proposes a visual detection method for motorcycle drivers wearing helmets. The proposed method is divided into two stages: motorcycle vehicle segmentation and classification (used to determine whether the target of interest is a motorcycle) and whether the helmet is worn or not. detection. For motorcycle classification, the commonly used Haar features are used as descriptors and SVM classifiers; for helmet detection, we use circular hough transform (CHT) to extract the head region, and then use the histogram of orientation gradient (HOG) descriptor to extract the image features, and use a multi-layer neural network (MLP) classifier to classify the target. The method has high detection accuracy, good real-time performance, and has engineering practical value.

A method for visually detecting the wearing condition of a motorcyclist's helmet, comprising the steps of:

Step 1) Acquisition of the background image: install the image acquisition device on the roadside to collect the road running conditions, set the actual road image extracted from the video stream as A, and use the adaptive Gaussian mixture model algorithm to extract it from the video stream The background image without moving objects is set to B;

Step 2) Segmentation of the moving target: segment the moving target from the captured image: first, subtract the image A and image B obtained in step 1) to obtain an image C; then, binarize the image C: use The otsu threshold segmentation method binarizes the image C to obtain the image D; finally, segment the moving object: perform edge detection on the image D, and use the morphological closure operator to de-dry the image, remove the image noise, and obtain the segmented The subgraph E representing the moving object;

Step 3) Target classification: divide the sub-image E obtained in step 2) into two types: motorcycle and non-motorcycle; first use Haar feature to map the detected object into a high-dimensional feature vector, and then use SVM classifier to judge Which category the image target belongs to; if it is judged to be a motorcycle, then enter the next step;

Step 4) Determine the region of interest and the head sub-window: first, the upper 1/6 to 1/4 part of the E image that is judged to be a motorcycle is defined as the region of interest, which is recorded as image G; then, the head Determination of the sub-window: use the circle Hough transform to calculate the circle in the image G, and record the circumscribed square corresponding to the sub-image with the best circularity in the image G as image I;

Step 5) feature extraction: use the HOG descriptor to perform feature extraction on the image I in step 4): wherein the HOG description is divided into nine pieces, and each piece is divided into nine small cells, so the resulting image consists of 81 sub-features A eigenvector H of

Step 6) Classification of sub-windows: After step 5) the image I in step 4) is subjected to feature extraction, a feature vector H in step 5) will be obtained, and this series of feature vectors H will be input to the multi-layer neural network The sub-windows are classified in the network MLP classifier, so that the images of the driver's head area are classified into two categories: those with a helmet and those without a helmet, and finally realize the detection of whether a motorcycle driver wears a helmet or not.

Further, the image acquisition device in step 1) is a CCD camera.

Further, the upper 1/5 of the E image of the motorcycle in the step 4) is the region of interest.

Description of drawings

Accompanying drawing 1 is the flow chart of embodiment of the present invention;

Beneficial effect:

1. The method is divided into two stages: motor vehicle segmentation and classification and detection of helmet use. For motorcycle classification, we use Haar feature as descriptor and SVM model as classifier; for helmet detection, we use circular Hough transform (CHT) for head region extraction, and use histogram of gradient orientation (HOG) descriptor to extract image features , and use a multi-layer neural network (MLP) classifier to classify the target. Through this classification method, the detection of whether the motorcycle driver wears the helmet is finally realized.

2. Through the two-level detection strategy of vehicle segmentation and classification and whether the helmet is used or not, the whole process of helmet detection is realized.

3. This method uses extensive road monitoring and visual sensing to solve the problem of not wearing a helmet detection due to its large amount of information and low cost.

detailed description

A method for visually detecting the wearing condition of a motorcyclist's helmet, comprising the steps of:

Step 1) Acquisition of the background image: install the image acquisition device on the roadside to collect the road running conditions, set the actual road image extracted from the video stream as A, and use the adaptive Gaussian mixture model algorithm to extract it from the video stream The background image without moving objects is set to B;

Step 2) Segmentation of the moving target: segment the moving target from the captured image: first, subtract the image A and image B obtained in step 1) to obtain an image C; then, binarize the image C: use The otsu threshold segmentation method binarizes the image C to obtain the image D; finally, segment the moving object: perform edge detection on the image D, and use the morphological closure operator to de-dry the image, remove the image noise, and obtain the segmented The subgraph E representing the moving object;

Step 3) Target classification: divide the sub-image E obtained in step 2) into two types: motorcycle and non-motorcycle; first use Haar feature to map the detected object into a high-dimensional feature vector, and then use SVM classifier to judge Which category the image target belongs to; if it is judged to be a motorcycle, then enter the next step;

Step 4) Determine the region of interest and the head sub-window: first, the upper 1/6 to 1/4 part of the E image that is judged to be a motorcycle is defined as the region of interest, which is recorded as image G; then, the head Determination of the sub-window: use the circle Hough transform to calculate the circle in the image G, and record the circumscribed square corresponding to the sub-image with the best circularity in the image G as image I;

Step 5) feature extraction: use the HOG descriptor to perform feature extraction on the image I in step 4): wherein the HOG description is divided into nine pieces, and each piece is divided into nine small cells, so the resulting image consists of 81 sub-features A eigenvector H of

Step 6) Classification of sub-windows: After step 5) the image I in step 4) is subjected to feature extraction, a feature vector H in step 5) will be obtained, and this series of feature vectors H will be input to the multi-layer neural network The sub-windows are classified in the network MLP classifier, so that the images of the driver's head area are classified into two categories: those with a helmet and those without a helmet, and finally realize the detection of whether a motorcycle driver wears a helmet or not.

Wherein, the image acquisition device in the step 1) is a CCD camera.

The upper 1/5 of the E image of the motorcycle in the step 4) is the region of interest.

specific embodiment

Step 1: Acquisition of background image: install the image acquisition device CCD camera on the side of the road, adjust the position and attitude of the image acquisition device to obtain high-quality images (video); set the actual road image extracted from the video frame Be A; Use the self-adaptive Gaussian mixture model algorithm (this algorithm is commonly used algorithm, do not add details) to extract the background image that does not contain moving object in the video stream and be B;

Step 2: Segmentation of the moving target: Segmenting the moving target from the captured image can be achieved through the following steps;

Step 2.1: Background difference method: Subtract images A and B obtained in step 1 to obtain image C:

Step 2.2: Image binarization. Use the Otsu threshold segmentation method (this algorithm is a commonly used algorithm, do not add details) to carry out binarization processing on the image C in step 2.1 to obtain image D: the selected threshold is automatically calculated by Otsu; use this threshold to image C. Binarization; pixels above the threshold are set to "1"; the rest of the pixels are set to "0";

Step 2.3: moving object segmentation: perform edge detection on the image D in step 2.2, and use the morphological closure operator to denoise the image to obtain the segmented sub-image E representing the moving object;

Step 3: Target classification: In the traffic scene, the objects segmented in step 2 (that is, the sub-image E) can be divided into two types: motorcycles and non-motorcycles; here, the classic Haar feature combined with the SVM classifier method is used : First use Haar feature to map the detected object to a high-dimensional feature vector; then use SVM classifier to judge which category the image object belongs to: if it is judged to be a motorcycle, then enter step 4;

Step 4: Determine the RoI and its sub-windows;

Step 4.1: Region of interest (RoI) extraction: the upper 1/5 part of the E image that is judged to be a motorcycle is defined as the region of interest, which is recorded as image G: this region of interest is obtained through the vehicle segmentation stage The experience value obtained by image statistics, the head area is usually located in the upper 1/5 image area;

Step 4.2: Determination of the header sub-window. Use the circle Hough transform CHT to calculate the possible circles in the image G; record the circumscribed square corresponding to the sub-image with the best circularity in the image G as image I, and this sub-window is regarded as the motorcycle driver's head area;

Step 5: Feature extraction: Feature extraction is performed on the image I in step 4.2 with the HOG descriptor. Among them, the HOG descriptor is divided into nine blocks, and each block is divided into nine small cells, thus generating a feature vector H composed of 81 sub-features;

Step 6: Classification of sub-windows: After feature extraction in step 5, each generated sub-window will get a feature vector, and this series of feature vectors H will be input into the multi-layer neural network MLP classifier to perform sub-window classification. Classification, the driver's head area image is classified into two categories with helmet and without helmet, so as to finally realize the detection of whether the motorcycle driver wears a helmet or not;

From step 1 to step 6, the whole process of helmet detection is realized through the two-level detection strategy of vehicle segmentation and classification and whether the helmet is used or not.

Adopt the method of the present invention, at first use C++ language to write motorbike driver's helmet detection software; Then video camera is installed on road side suitable position, and vehicle image is collected during vehicle running; Subsequently, the original image that takes is input It is processed in the helmet detection software; the resolution of the video is 1280*720 pixels and 30 frames per second, the total time of the video is 150 minutes, and the accuracy rate of the vehicle classification results reaches 97.78%. The helmet detection algorithm achieved an accuracy rate of 91.37%. The operating environment is WinXP, and the CPU is 2.4GHz.

In summary, the present invention takes vehicle segmentation and classification and the detection of helmet use as a complete system, fully utilizes the Haar feature descriptor and the SVM model as the high precision obtained by the classifier in the vehicle classification stage and uses the circular Hough transform (CHT ) to accurately extract the head region, and then use the histogram of gradient orientation (HOG) descriptor to extract image features, and use the multi-layer neural network (MLP) classifier to classify the target. The characteristics of the high evaluation index obtained in the helmet detection stage, Therefore, a method for accurately detecting whether the driver is wearing a helmet from the provided input source image is realized.

The described embodiment is a preferred implementation of the present invention, but the present invention is not limited to the above-mentioned implementation, without departing from the essence of the present invention, any obvious improvement, replacement or modification that those skilled in the art can make Modifications all belong to the protection scope of the present invention.

Claims (3)

1. the visible detection method to helmet of motorcycle driver wear condition, it is characterised in that comprise the steps:

Step 1) acquisition of background image: image collecting device is installed on roadside place road operation conditions is acquired, the road real image extracted from video flowing is set to A, and the background image extracted in video streaming without moving object with adaptive GMM algorithm is set to B；

Step 2) segmentation of moving target: moving target is split from the image of shooting: first, by step 1) in the image A and image B of gained subtract each other, obtain image C；Then, to image C binaryzation: by otsu threshold segmentation method, image C is carried out binary conversion treatment, obtain image D；Finally, to moving meshes: image D is carried out rim detection, and closes operator to image except dry with morphology, remove picture noise, obtain the subgraph E of the representative moving object being partitioned into；

Step 3) target classification: by step 2) in the subimage E that obtains be divided into two kinds: motorcycle and non-motorcycle；First by Haar feature, detected object is mapped as a high dimensional feature vector, judges which kind of image object belongs to followed by SVM classifier；If being judged as YES motorcycle, then enter next step；

Step 4) determine area-of-interest and head subwindow: first, will be deemed as being that upper 1/6～1/4 part of the E image of motorcycle is defined as area-of-interest, be designated as image G；Then, the determination of head subwindow: calculate the circle in image G with Circle Hough Transform, the circumscribed square having in image G corresponding to the subimage of best circularity is designated as image I；

Step 5) feature extraction: with HOG descriptor to step 4) in image I carry out feature extraction: wherein HOG describe be divided into nine pieces, every piece is divided into nine junior unit lattice, then creates the characteristic vector H being made up of 81 subcharacters；

Step 6) classification of subwindow: through step 5) by step 4) in image I carry out feature extraction after, capital obtains a step 5) in characteristic vector H, being input in multilayer neural network MLP grader by this series of characteristic vector H child windows classifies, thus the head zone image of driver has been categorized as the helmet and without the big class of the helmet two, thus finally achieving the detection whether helmet of motorcycle driver is worn.

2. a kind of visible detection method to helmet of motorcycle driver wear condition according to claim 1, it is characterised in that described step 1) in image collector be set to ccd video camera.

3. a kind of visible detection method to helmet of motorcycle driver wear condition according to claim 1, it is characterised in that described step 4) in motorcycle E image upper 1/5 be area-of-interest.