Multi-scene Human Detection And Counting Method Based On Feature Learning

Human detection and counting is an important branch in pattern recognition,and has important applications in security,intelligent interaction and so on.In the relatively sparse scene with individual monitoring needs,the detection algorithm based on individual characteristics is usually adopted.In the relatively dense scene with monitoring the group security state needs,the detection algorithm based on global features is usually adopted.Due to the complexity of the background and the variety of human body's non-rigidity,and other external interference,how to ensure the accuracy of the algorithm and reduce the computational complexity and enhance scalability are the urgent technical problems in the above fields.The paper briefly introduces the application scenes and recent study in the above two scenarios,mainly analyzing the mainstream algorithm framework and existing problems.To solve the problem of false-target interference and multi-scale repetitive detection in the former scene,an optimized aggregated channel features pedestrian detection algorithm based on binocular vision is proposed.Binocular vision improves the accuracy of target segmentation,and a two-step segmentation strategy is proposed to adaptively optimize the pyramid hierarchy structure.To solve the problem of imprecise foreground segmentation and poor scalability of model trained in the latter scene,a population counting algorithm based on dynamic corner and local clustering is proposed.Moving feature points are clustered based on density,and the effective occlusion factors are extracted to train a more expansive classification model.Simulation results show that the former optimization algorithm compensates for the increase of computational complexity caused by binocular information,and saves nearly 60% of running time compared with the control method and effectively solves the problem of multi-scale repetitive detection.The latter optimization algorithm can achieve more precise counting result with only 3-D features,with MSE of 2.357 and MAE of 1.093,which outperforms the state of the art approaches.