Research And Application Of Real-time Target Detection Based On Deep Convolutional Neural Network

Pedestrian fine-grained detection and pedestrian multi-attribute recognition provide useful clues for intelligent video analysis,arouse people's interest in intelligent video surveillance analysis and pedestrian re-identification,and play a vital role in intelligent video surveillance and security-centric computer vision systems.Pedestrian fine-grained detection locates component information in static or dynamic pedestrian images,such as head,upper body,lower body,shoes,bag,and hat.Pedestrian multi-attribute recognition can identify pedestrian attribute information,such as gender,type of hair,upper-body clothing,lower-body clothing,type of shoe and action.Convolutional neural networks(CNN)can automatically learn the features of images,which is widely used in computer vision tasks.But in complex surveillance scenarios,it is a challenging task for CNN to learn the fine-grained features of pedestrian components and pedestrian attributes.Therefore,the research and application of real-time object detection based on deep convolutional neural network has extensive research value and application prospects,especially for pedestrian fine-grained detection and pedestrian multi-attribute recognition.The research content of this paper is pedestrian fine-grained detection and pedestrian multi-attribute recognition,which are mainly divided into the following aspects:(1)As for the difficult problem of pedestrian fine-grained detection,we propose a novel deep convolutional neural network(DCNN),called MSF-ACNN.The framework uses atrous convolutions in cascade and multi-scale feature fusion to improve the accuracy of small object detection.Atrous convolution effectively expands the field-of-view of the convolution kernel and preserves the context information of the object to improve the accuracy of the small object without increasing the number of parameters and computation.Multi-scale feature fusion obtains more meaningful fine-grained information from both the low-level and high-level feature maps,and can handle a variety of image scales.The experimental results show that MSF-ACNN can obtain better mean average precision(m AP)than the current state-of-the-art methods with faster detection speed,achieving significant improvements on certain small parts of pedestrian such as shoes,bag and hat.(2)In order to accurately identify the fine-grained attributes of pedestrians,we propose a DCNN-based MSE-Net framework.The framework uses the MLSC module to obtains more meaningful fine-grained features in low-level and high-level features,and retains contextual information on the characteristics of pedestrian fine-grained attributes,such as glasses and accessories.The SE-block module enhances network sensitivity to information and compresses features with a global receptive field.It can automatically acquire the importance of each feature channel through learning.Finally,the useful features are enhanced according to this importance level and the features that are not useful for the current task are suppressed.The experimental results show that the MSE-Net framework outperforms the state-of-the-art methods on RAP(Richly Annotated Pedestrian)dataset,and the robustness against predicting positive and negative samples in each attribute.(3)In order to simultaneously achieve the task of pedestrian fine-grained detection and multi-attribute recognition,we propose a DCNN framework based on multi-task learning for pedestrian analysis,with the main idea of integrating different learning tasks of pedestrian body parts detection and pedestrian attribute classification,which called Hyper-pedestrian convolutional neural network(HP-CNN).The HP-CNN framework mainly includes SE-block modules and multi-scale feature fusion modules.SE-block strengthens the representational power of networks by selectively emphasizing informative features.Multi-scale feature fusion concatenates more fine-grained information from both the low-level and high-level and enhances the contextual information from different convolutional layers.We validated and evaluated HP-CNN on the authoritative RAP dataset.The experimental results show that HP-CNN obtained better results of both body parts detection and multi-attribute classification.The body part detection task and the multi-attribute recognition task cooperate with each other to significantly improve the performance of multi-attribute recognition,effectively balancing the speed and accuracy of body part detection.