Research On Pedestrian Detection Technology Based On Deep Learning

In recent years,pedestrian detection technology has played an indispensable role in video surveillance,automotive driverless technology,intelligent robots and other fields.And it has become a very popular research topic in the field of modern computer vision.Computer vision based on image recognition technology has become a very important research topic of artificial intelligence technology.How the computer understands the scene more humanely and recognizes the category of each object in the scene like a human is the key to solving the problem of the application of the artificial intelligence system.Compared with traditional image recognition systems,the biggest advantage of deep neural networks lies in the integration of feature extraction and classification of images into an overall neural network.At the same time,with its deepening network structure,it is possible to extract more advanced features that are more abstract than human hand-designed features.This paper focuses on the pedestrian detection and recognition problems of robots and unmanned vehicles in natural scenes.The research on pedestrian detection algorithms based on deep learning is mainly carried out.The main research content is as follows:(1)Propose a pedestrian detection method based on fuzzy deep learning network.Based on the classical deep belief network,combined with the idea of fuzzy sets,the deep learning framework embeds prior knowledge and restricts the Boltzmann machine.The combination of abstract ability and excellent classification performance of fuzzy sets was successfully applied to pedestrian detection.Through the proposed fuzzy deep learning network,a higher level of abstract features of descendants in complex backgrounds can be extracted,the overall performance of the system can be improved,and the detection accuracy of the system can be further improved.(2)A dynamic adaptive pooling model is proposed.Compared with the traditional average pooling and maximum pooling,the adaptive pooling presented in this paper is more specific for different image feature extraction and can be based on specific features.The graph dynamically adapts the process of adjusting the pooling operation.At the same time,the corresponding weights can also be adaptively adjusted according to the specific content of the internal data of each pooled domain.This also improves the processing capabilities of the deep learning network for different pooled domains as a whole.The neural network can extract more effective features during different iterations and processing different pooled domains.(3)A fast pedestrian detection model based on dynamic adaptive regional convolutional neural network is implemented.Based on the classical fast regional convolutional neural network,a dynamic adaptive pooling algorithm is incorporated,and the maximum pooling in the classic network is used as an adaptive pooling,making the new network further improve the processing performance of pedestrian features.The promotion can more specifically deal with different pedestrian characteristics data.Compared with the traditional fast area network,the training speed of the neural network is further accelerated and the detection accuracy of the pedestrian detection model is improved,and at the same time,the robustness of the system and the real-time performance of the monitoring are also improved.