On Distributed Compressive Sensing And Profile Recognition

Effective monitoring of the enclosures at all times, such as detecting smugglers and illegal immigrants on the border, has great significance to public safety. The video surveillance system can get a better video effect, but huge amounts of data challenge practicality of intelligent detection and recognition.Across the border or uninhabited sections, Sensing information has great pertinence, only concerned about whether someone entry or exit, and do not need to identify the animalsand leaves swing. Accordingly, the use of active photodetection profile feature to achieve automatic monitoring is a method of feasibility. The system includes emission and collection of the active light source, compression sampling, intelligent detection and recognition; simultaneity, It also involve other problems about constituting the wireless monitoring network: such as distributed data fusion, reducing network transmission power and so on.Based on the effective monitoring of the enclosures at all times, both profile feature and recognition algorithm were studied. The problems of data acquisition, fusion and specific object recognition in multi-class sample were focused in wireless monitoring network. The main works and contributions of the thesis were outlined as follows:1) Combined with the theory of sparse representation, a novel method of constructing a overcomplete dictionary is proposed. When objects pass through sensors field of view of profile detecting system, the velocity, posture, angle and other external factors are different, the dimension of the object profile feature samples captured by the sensing unit is also different. For ease of processing, the thesis pretreated the profile samples by principal component analysis, which extracted main components of the sample signal to eliminate redundant information, projected lower dimensional space to construct a dictionary and obtained the eigenvectors to construct a dictionary by transforming the samples into the same size of the feature vectors.2) Based on the above dictionary, a novel algorithm of special object recognition based on sparse representation is proposed (that is SRSOR). The algorithm is based on the minimum residual to determine the category of test samples, and compare with traditional recognition algorithm from multiple aspects. Experimental results demonstrate the effectiveness of the algorithm.3) Combined with the theory of wireless sensor network and distributed compressive sensing, a method of data fusion based on distributed compressed sensing is proposed. If a signal has a sparse representation in one basis, according to compressive sensing theory, the original signal can be exactly recovered from a small number of measurements, mainly taking use of intra-signal structures at a single sensor. However, for densely distributed wireless sensor network, signal correlation may exist among the sensor nodes. The theory of distributed compressive sensing enabled new joint coding algorithms that exploited both intra-and inter-signal correlation structures. On this basis, measurements of sensor notes were fused by accumulate mode to further reduce the amount of data and a mathematical model was proposed.4) Based on the above data fusion method and its mathematical model, an algorithm of special object recognition in multi-class fusion sample is proposed based on morphological component analysis. In order to recognize the classes contained in the fusion sample, fusion sample was sparsely represented by over-complete dictionary contained more than one class; then, according to distribution of the main non-zero coefficient in the sparse representation coefficients, the algorithm recognized the classes included in fusion sample, and further judged whether there was a specific object. The experimental simulation results showed the effectiveness of the specific object recognition algorithm based on morphological component analysis in some aspects.5) Designed and implemented a hardware test platform. The core device of the platform is FPGA chip, sensing devices is the active reflective photoelectric sensor with the stable performance; those sensors was assembled at uniform intervals in the vertical holder about2m; When the object passed through the sensors field of view, the control unit read and stored the sensor group output state in parallel until the object was completely out of all sensor FOV. Considered state data of the sensor group as information of passing object’ profile feature, and the data was processed in future. The hardware structure of test platform was simple and easy to upgrade. According to actual needs, we could increase the number of sensors in the sensing unit, and a patent has been granted.