Research On Pedestrian Detection Technology For Real Scene

With the rise of artificial intelligence and continuous improvement of computing power,deep learning algorithms based on neural networks have shown amazing results.Especially in terms of visual perception,computers have surpassed humans in many tasks.Therefore,this thesis begins with the pedestrian detection of this particular visual task and improves the pedestrian detection algorithm.The main work of this thesis is as follows:(1)Introduce the design principles and guidelines of convolutional neural networks,introduce the basic components of convolutional neural networks,loss functions,regularization methods,network optimization and acceleration methods,and its specific applications.(2)The pedestrian detection method based on Faster R-CNN is mainly studied.Of the current target detection algorithms based on deep convolutional network,Faster RCNN is one of the most classical methods.Due to the similarity of pedestrian detection and generic target detection,the Faster RCNN can be used to detect pedestrians.However,if pedestrians and general targets,such as vehicles,differ in terms of form and other aspects,if the Faster RCNN does not make corresponding customized corrections according to pedestrian characteristics,it is bound to fail to achieve optimal performance.So this thesis first studied how to apply the Faster RCNN to pedestrian detection problems.(3)Propose a coarse-to-fine pedestrian detection method.As its name implies,it is completed by two stages of testing.Firstly,a fast and high-performance pedestrian detector is used to complete the rough extraction of pedestrian candidate boxes.A coarse classification network is used to subdivide the rough extraction structure on the result of the rough extraction in the previous step,thereby completing the entire pedestrian detection process.(4)Proposed a method to use semantic segmentation information to change the human detection performance.Pedestrian detection is the main problem of computer vision.It mainly includes two subtasks,one is detection and the other is semantic segmentation.The two tasks have certain similarities and differences.Pedestrian detection can give different pedestrians' bounding boxes,but pedestrians' boundariescannot be given.Semantic segmentation can accurately give pedestrians boundaries,but it is not easy to distinguish between different objects.This thesis discusses how to use semantics to segment information without affecting the efficiency of pedestrian detection.Improve the accuracy of pedestrian detection.And propose a network structure to experiment with this idea.