A method for detecting miners based on helmets detection in underground coal mine videos

Mining Science and Technology( China)21(2011)553-556Contents lists available at Science Direct氵Mining Science and Technology( China)ELSEVIERjournalhomepagewww.elsevier.com/locate/mstcA method for detecting miners based on helmets detection in undergroundcoal mine videosCai Limei" Qian JianshengSchool of Information and Electrical Engineering. China University of Mining 8 Technology, Xuzhou 221116, ChinaARTICLE INFOA BSTRACTIn order to monitor dangerous areas in coal mines automatically, we propose to detect helmets fromReceived in revised form 18 January 2011underground coal mine videos for detecting miners. This method can overcome the impact of similarityAccepted 16 February 2011between the targets and their background. We constructed standard images of helmets, extracted fourAvailable online 10 August 2011directional features, modeled the distribution of these features using a Gaussian function and separatedlocal images of frames into helmet and non-helmet classes, Out experimental results show that thismethod can detect helmets effectively. The detection rate was 83.7%Human detectiono 2011 Published by Elsevier B V on behalf of China University of Mining TechnologyHelmet detectionmage pattern recognition1 Introductionwith videos. This problem is difficult to deal with when the targetsare occluded or similar to the background, conditions prevalent inThere are several dangerous areas in coal mines, such as areas underground coal mines [31with harmful gas, inclined laneways in which winches are moving The characteristics of coal mine videos are affected by all-or equipment rooms, where miners are not allowed. But the lane- weather artificial illumination and the conditions of undergroundways crisscross and are like a maze. warnings in the form of signs laneways, such as: (1)low illumination: artificial illumination difor alarm bells are insufficient. It is important to monitor these fers from the natural light and is clearly low(2)uneven lightingareas effectively. Infrared methods to detect the direction and where illumination is strong near the object to be illuminatednumber of miners are available [1]. But, the technique of identifi- but obviously insufficient far away from it, hence the outline ofcation and human positioning must also be used if we want to an object is indistinct (3)a lack of color information in imageknow the identity and location of miners [2 ). The application of a processing: there are only black, white or gray tones in coal minesmonitoring system is restricted because of its complexity. These except for some equipment; (4)the color of mine clothing is darkdays, most coal mines have been equipped with video monitor sys- blue or dark gray which is similar to the background in low illumi-tems. If miners can be detected using videos and the system can nation. Sometimes, we need to pay more attention to the videos forsound the alarm on time and control linkages automatically. many recognizing miners: (5)cameras are usually installed on lanewayaccidents can be avoided. At the same time, we can obtain much walls, often only a little higher than the height of a person, whereuseful and important information for safety in production. There videos of miners usually do not show them whole-length orare many methods for detection of people which in general can upright, hence the methods described by various authors cannotbe divided into two categories: approaches based on motion detec- be used directly [4-8, 11-14]tion such as background subtraction, frame difference or opticahese conditions make it more difficult to detect miners in coaflow techniques, and approaches that detect people directly from mine videos and cause the results of traditional detection methodsstatic images (3]. The second category approaches are designed to be unsatisfactoryto classify and distinguish people from inanimate objects [4-8)We propose a method to detect the helmet of miners. MinersThese approaches can also be used with videos [9, 10]. Although are obliged to wear a helmet when working underground whichthey have been studied for a long time, there are no robust and fast because of their material, reflect light. the upper parts of a helmetmoving target detection methods because of various interferences in images are usually different from their background. If a helmetding author. Tel. +86 13775892667Our presentection 2 our methodE-mailaddress:Imcai1977@163.com(LCai).to detect mine中国煤化工ribed and analyzedCNMHG1674-5264/s-see front matter 0 2011 Published by Elsevier B V, on behalf of China University of Miningdoi:10.1016 J.mstc.2011.06016L Cai, J. Qian/Mining Science and Technology( China)21(2011)553-556△o●DQo●画·画A(a)Helmet images(b) Non-helmet imagesFig. 1. Examples of images used in training.Section 3 shows our experimental results and we present our conclusions in Section 4PreprocessingSubtract2. Method to detect minersFilterocal imagWe constructed standard images of helmets, extracted fourdirectional features. modeled the distribution of these featuresusing a Gaussian function, designed the classifier, and separated lo-Extraccal images of frames into a helmet and a non-helmet class fromOutputfeatureswhich we detected the miners2.1. Collecting training samplesFig 3. Detection flowchart.The helmets used in coal mines are mainly of two shapes: those non-helmet training data In the end, there were 961 non-helmetwith a small visor and those without. We constructed two kinds ofhelmet images from different viewpoints. The helmet was inclined Images in the training sample set. Some images used in trainingwere0°,±10°,±20°,±30°and±90. The vertical directions of thesepoints changed from 0o to 90 by 3. These directions contained all 2.2. Extracting featurespossible situations in the actual environment. We constructed2100 standard helmet images in all. They were normalizedIllumination is always placed on the top of laneways. the uppercoal mine videos using the bootstrap method as described by Kah the helmet is usually different from the background in the videosand Poggio[15]. Hence, there are 2669 images in the training sam- It means that the upper part of edges is at least clear. We selectedthe four edge images as a vector.Bootstrapping is an iterated procedure to collect training dataBy applying Prewitt's operator in four directions, i.e. vertically,for classification. It starts with a small learning sample set for horizontally and along both diagonals, four edge images were con-training the initial classifier, adds misclassified data to the training structed from a helmet image. Each edge image was filtered usingdata, and then trains the classifier. this is repeatedGaussian filters and resized to 8 x 8 pixels. The near-boundary pixelsand then the edge imagto collect non-helmet samples efficiently. Generally, the non- circle with a diameter of 6 pixels. The four circles made a featurehelmet areas are any areas, subject to video monitoring, where vector with 96 dimensions. Two standard helmet images, a realhelmets are not worn. Considering computational costs, a back- helmet image and their four edge images are shown in Fig. 2ground subtraction would be used to establish scanning arewhere miners are present, areas lit by miner's lamps and some 2. 3. Modeling the distribution of helmet and non- helmet pattemsareas where noise is heard. the shoulders of miners and areas litby their lamps are curved so that these areas will interfere withThe two kinds of helmet have about 10 shapes( Fig. 1), takenclassification. We selected 200 images of this kind of area as the from different points of view, hence we modeled the distributionnon-helmet images, and then used the bootstrap method to collect of helmet patterns using 10 Gaussian functions中国煤化工CNMHGFig 2. Helmet images and their four edge images.L Cai J Qian/Mining Science and Technology(China)21(2011)553-556(a)Origin frame(b) Binary difference imageg regionsFig. 4. Result of preprocessing.(a) Helmet regions detected(b)Segmentation results of absolute difference images using OtsuFig 5. Detection results.Non-helmet areas mainly include shoulders and areas lit byWe translated this absolute difference image into a binary im-lamps. We modeled the distribution of helmet patterms using three age in order to simplify the operation and to eliminate smallGaussian functions, k-means algorithm was used to cluster the noises. The threshold value cannot be too large to preserve the ob-vectors based on Mahalanobis distances. The Mahalanobis distance ject. When the difference is more than 10 gray-scales, two gray lev-between the test pattem x and the cluster centroid u is given by: els can be clearly identified. So, we chose 20 as the threshold valueTin our expM(x,)=(x-)2x1(x-p)(1)J(xy)-(xy)≥TThe following is an outline of our k-means clustering procedure: BW1(x,y)=0 U(x,y)-So(x,y)