Pedestrian And Vehicle Detection Using Improved YOLOv3 Network With Multiscale Feature Fusion

Pedestrian and vehicle detection is one of the research directions in computer vision,and is widely used in video surveillance,intelligent robots and other fields.Aiming at the detection of pedestrians and vehicles in urban scenes,this paper focuses on the multi-scale feature extraction of convolutional neural network and Bounding Box(BBox)regression,and proposes a improved YOLOv3 network method for pedestrian and vehicle detection.In order to detect objects of different scales,In this work,by making full use of the multi-scale features of convolutional neural network(CNN),we propose a multi-scale feature fused YOLOv3 network to improve the detection accuracy of multiple scale objects in images.Firstly,based on the YOLOv3 network,we adopt the Scale-Transfer Module of STDN algorithm to narrow down the lowlevel feature map,and construct a feature-reused backbone network for extracting features.Then,the the Scale-Transfer Module is used to amplify the high-level feature maps and thus construct a Feature Pyramid Network(FPN)for prediction objects.On the public MSCOCO detection challenges,the network can effectively improve the false and missed detection problems caused by the occlusion and small object of the original YOLOv3 network,so the accuracy of the experimental results of each category object is better.For the two tasks of object detection and classification,we propose a detection model based on confidence optimization and cascaded BBox regression.Inspired by the two-stage object detection model Cascade R-CNN,this paper uses the cascaded BBox regression method to modify the onestage object detection model improving the accuracy of object detection.This architecture sets the incremental IOU(intersection over union)threshold to regress BBoxes at training,which obviously improves the quality mismatch of optimal IOU value,and detection deviation problem of BBox.At the same time,the classes confidences is optimized according to the regularity characteristics of the object attribute of the classes,so as to reduce the false detection that violate the statistical law and the error conditions such as the correct detection box being filtered out due to the low confidence.This method is based on the prior law that the object categories are always interdependent in the real scene.Firstly,we constructe the correlation coefficient matrix and the exponential normalized inter-class object attribute regularity,and then the confidence of the object is optimized in the detection stage.This confidence optimization method utilize the objective existence rule as the information to be applied to the object detection,and the object detection effect is slightly improved.Aiming at the problem of pedestrian and vehicle detection in urban scenes,In this work,we proposes a pedestrian and vehicle detection model based on multi-scale feature fused YOLOv3 network and cascaded BBox regression.Firstly,considering the factors of occlusion,light,scene and so on which bring negative effect on detection,we constructe an effective pedestrian and vehicle data set,and use the sample imbalance method to solve the sample imbalance problem.Then,in order to speed up the convergence of the model at training and reduce the probability of missed detection,the K-means clustering optimization method is used to select the a priori box.Finally,we adopt the multiscale feature fused YOLOv3 network and the cascaded BBox regression model for pedestrian and vehicle training and detection.The experimental results show that the pedestrian and vehicle detection methods based on the improved YOLOv3 network are feasible and effective,and have far-reaching reference value.