Key Technology Research Of People Counting Under The Surveillance Environment

People counting has been widely used in the fields of crowd surveillance, intelligent transportation and so on. With the popularity of video surveillance technology, how to count the number of people under the surveillance environment quickly and accurately has been an important part in the fields of image processing and video surveillance. But the task is facing great challenges, such as heavy occlusion, low resolution, imaging viewpoint variability, etc. To solve these problems, this thesis mainly studies key technologies of people counting under the surveillance environment, including work in the following two aspects:First, for the scratches problem of Block Matching and 3D filtering(BM3D) method, this thesis presents an image denoising method based on the methods of BM3 D and guide filtering. The method reduces the phenomenon of obviously scratches in the smooth area which are produced by the BM3 D method by combining the methods of BM3 D and guide filtering and utilizing the strategy of second joint filtering. Specifically, firstly, the initial filtered image is got by BM3 D denoising method; Secondly, smoother guide image is got by BM3 D denoising in low scales, and the initial filtered image is filtered again with the guidance of the guide image. The results show that the proposed algorithm can not only maintain the excellent denoising performance of BM3 D algorithms, but also has better visual effects and higher performances.Second, inspired by the successful application of Region Convolutional Neural Network(R-CNN) on object detection, the thesis proposes a people counting method based on head detection which combine the algorithms of Adaboost and the Convolutional Neural Network(CNN). Unlike the R-CNN which used the general object proposals as the input of CNN, this method uses the cascade Adaboost algorithm to obtain the head region proposals as the input of CNN, which can greatly reduce the following classification time. In view of the strong feature learning ability of CNN, it is used as a feature extractor in the thesis, instead of as a classifier. The final classification is done by a linear support vector machine(SVM) classifier which train the features extracted by CNN feature extractor. Finally, this thesis applies the prior knowledge to post-process the detection results to eliminate the false detection target and get the final results of people counting. In order to evaluate the proposed method, the thesis tests multiple images from classrooms under the surveillance environment and PETS2009 images under the outdoor environment respectively. The test results show that the method can detect the head of the test image accurately and outperform the classical methods of deformable part model(DPM) and cascade Adaboost. Currently, the people counting algorithm of this thesis has been used for surveillance system of teaching building in the school.