| Coal is the primary energy of our country and coal mine safety production is of great significance. Enhancing the level of automation and informatization in coal mine production and establishing safety monitoring system, personnel location system underground meet requirements of six main safety protection systems' constructions, which are also the important safeguards for mine safety production and emergency rescue. The wide use and development of digital mine video system lays the hardware foundation for realizing intelligent video surveillance system and miner visual tracking location system. The applications of computer vision method in coal mine personnel location system and safety monitoring system have been investigated to further optimize image intelligent recognition and tracking theory and promote intelligent level underground in this thesis. Considering coal mine characteristics, target especially miner tracking, recognition and coal-rock image recognition in coal mine video surveillance images are mainly investigated. Because of the particularity in coal mine, the wireless transmission is necessary for video real-time transmission in coal mine video surveillance system. But in extreme conditions, energy carried by radio frequency electromagnetic wave can lead to spark, which may ignite gas. The safety distance between radio sources and energy coupling structures should be considered to deploy wireless video image acquisition nodes. Thus, safety of electromagnetic wave energy has been discussed in appendix.Personnel uniqueness detection technologies requirements are discussed considering the mine environments, which is an important part of personnel tracking and positioning system. We also discuss the merit and demerit of different biometric identification methods in uniqueness detection. The face recognition method and system based on image morphology in thermal infrared image is proposed. The blood vessels net of human face is obtained through anisotropic diffusion filter and image morphology. Then the mending and skeleton extraction methods for vessel net have been proposed. Based on the skeleton image and the intersection points of vessel, the distribution vector of intersection points is computed to describe and recognize the human face. Because the temperature of miner's lamp suface increases and glasses will block the transmission of infrared ray, these factors will affact the extraction of intersection points. So the detection methods for miner's lamp and glasses are proposed. Besides, thermal infrared image is sensitive to the environment temperatures, which will affact the recognition rate. A temperature compensation method is proposed to address this problem. Experiments show the reliability and recognition results of this algorithm.The basic characteristic and evaluation methods of miner visual tracking are discussed. To address the problems of real-time and precision requirement, illumination variation and occlusion in coal mine worker visual tracking and locating, a real-time multi-scale target tracking algorithm based on compressive sensing is proposed by fusing the scale invariant random projection appearance model into the bootstrap filter framework, which is an approximate approach for bayesian dynamic state estimation. The algorithm is robust to occlusion, illumination and scale change. Scale invariant normalization rectangle feature is designed to describe the appearance of target with different scales to track. The scale invariant property for normalization rectangle feature is proved. Experiments represent an effective, simple and robust method for scale variation or in-variation targets. These normalization rectangle features of different sizes and coordinates are effective and efficient computed with the technique of integral image method. To compresive capture lower dimension compresive feautures with different scales, scale adaptive random matrix is designed according to the theory of compressive sensing, which achieves a real-time tracking. The observation model of bootstrap filter is constructed with the response value of na?ve bayes classifier. And the importance weights are used to re-sample the particles and avoid degeneracy of samples. Two feedback strategies are proposed to cope with the short time and long time occlusion. The experimental results from various benchmark challenging sequences demonstrated the superior performance in precision, stability, speed and adaptation of our tracker when compared with some other state-of-the-art tracking algorithms.In order to realize the real-time tracking based on mean-shift, a fast mean-shift iteration algorithm is firstly proposed. The factors limiting mean-shift iteration convergence rate has been analyzed. An extension method of mean shift vector iteration steps is proposed. The astringency of the impoved method is proved and the condition of convergence is deduced. The iteration pathes and the possible oscillation stage are further discussed. We also analyze the influence of oscillation to convergence rate and use the acceleration iteration theory to optimize the oscillation stage. By testing the rate of convergence of basic mean-shift iteration function, the feasibility of adopting acceleration iteration theory is proved. Experiments show the effects to convergence rate with different scaling factors and verify the acceleration for oscillation stage.Based on the fast mean-shift method, a miner visual tracking method is realized. Compressive feature with kenel and Bhattacharyya coefficient is used to construct the similarity measurement between target and candidate target models. The real-time iteration search of Bhattacharyya coefficient maximization is realized by adopting the proposed fast mean-shift method. Aiming at the problems of the influence of miner light carried by miner on target appearance, especially in human visual tracking underground, a real-time multi-scale multi-feature fusion method is proposed. To relief the effect of miner light for target appearance, the confidence of sample was computed using marginal color feature and was fused with scale invariant compressive feature appearance model to formulate the observation model of particle filter framework. A second-order model is chosed as the motion model of bootstrap filter, which considers the prior velocity information and makes the sampling more effective. Experimental results show the superior performance of our tracker compared with several state-of-the-art trackers in terms of precision, stability and adaptation to special environment underground. Coal mine computer vision personnel tracking system structures accord with the characteristic of coal mine are discussed in this thesis, including whole mine miner visual tracking and location system and target behavior analysis alarm system in key areas and restricted areas. Deep mining is the main mode in coal mine of our country. The tunnel is long and narrow, full of branches and corners. So the system is suitable for multiple cooperating cameras tracking by adopting tree branches structures plus relay manner. The camera arrangements and relay tracking solutions are designed for long straight tunnel, three-bifurcation tunnel, bend tunnel, cross tunnel, annulus special turnnel, triangle special turnnel, et al.Coal-rock recognition technology based on image analysis is studied for the key technology in working face manless mining by using the distinction between captured coal and rock sample images in gray level and texture. A method of texture analysis in multi-resolusion based on multiwavelet transform has been proposed and a coal-rock recognition system based on image is designed.Considering system maintenance or bandwidth and distance requirements, because of wicked environment underground, broadband wireless transmission technology is necessary in mine intelligent video surveillance system. In worst cases, radio frequency electromagnetic may ignite gas. The safety of electromagnetic wave energy on gas was evaluated with two types of electromagnetic wave energy coupling in appendix, including coupled magnetic resonance and electromagnetic radiation resonance. The safety distance formulas between RF sources and the coupling structures were presented for different coupling types. For electromagnetic far field radiation, the ignition mechanism and the maximum power dissipated in load from metal structures which act as electrically-small or electrically-large receiving antenna was investigated. Based on coupled magnetic resonance, the maximum energy transmission efficiency was studied. Combining the feature of mine, the safety distances for different frequency bands are obtained, which provides the basis for wireless video acquisition nodes' deployment.The innovations in this thesis lie in six aspects as below:(1) The personnel uniqueness detection method and system based on miner biometric features of human face in thermal infrared image are developed and designed according to the requirements and indexs of uniqueness detection system specified in coal mine personnel location system. The detection and processing methods of miner's lamp and glasses are proposed. The temperature compensation method is developed to address the influence of environment temperature discrepancy.(2) The real-time multi-scale target visual tracking algorithm based on compressive sensing is proposed to address the requirements of precision, speed or real-time property, illumination variation, target scale variation and occlusion of coal mine visual tracking.(3) Mean-shift based tracking is a classical deterministic tracking method, but the low convergence rate of mean-shift limits the sppeed of tracking. A fast mean-shift iterative method is proposed to improve the speed of convergence of the method. An extension method of mean shift vector iteration steps is proposed. The astringency of the impoved method is proved and the condition of convergence is deduced. The influence of oscillation to convergence rate if analyzed, and the acceleration iteration theory is used to optimize the oscillation.(4) Human target matching and tracking method in coal mine video is proposed to meet the coal mine condition, including a real-time multi-scale multi-feature fusion miner visual tracking method and miner visual tracking method based on compressive feature with kenel and the fast mean-shift method. Aiming at the problems of the influence of miner light carried by miner on target appearance, especially in human visual tracking underground, the confidence of sample was computed using marginal color feature and was fused with scale invariant compressive feature appearance model to formulate the observation model of particle filter framework. Miner visual tracking method based on fast mean-shift iteration and compressive feature with kenel is also realized. Based on the fast mean-shift method, compressive feature with kenel is designed and Bhattacharyya coefficient is used to construct the target and candidate target models. Coal mine computer vision personnel tracking system networking mode and structures are proposed accord with the characteristic of coal mine, considering the long and narrow tunnel, which is full of branches and corners.(5) A method of texture analysis in multi-resolusion based on multiwavelet transform has been proposed. The texture energy values in dfferent resolusions are computed from the transformed domain. Multi-resolusion texture energy distribution vector is obtained based on these texture energy values. Texture could reflect gray relations between neighbour pixels and will be seldom affected by illumination. A coal-rock recognition system based on image is designed.(6) The safety of electromagnetic wave energy on gas was investigated with two types of electromagnetic wave energy coupling, including coupled magnetic resonance and electromagnetic radiation resonance. The coupled magnetic resonance model and electromagnetic radiation resonance model are established. The safety distance formulas between RF sources and the coupling structures were presented for different coupling types. Combining the feature of mine, the safety distances for different frequency bands are obtained. |
PDF Full Text Request