Multi-branch target detection method based on traffic scene

Abstract

The invention requests to protect a multi-branch target detection method based on a traffic scene, and the method comprises the steps: S1, obtaining a high-definition picture taken by a traffic intersection to construct a related data set, carrying out the classification and marking of traffic scene images, generating a corresponding category label, and dividing a training set and a test set; s2,building a network model with 32 layers based on deep learning, obtaining nine anchor frame priori through a k-means clustering algorithm; averagely distributing the nine anchor frames into three detection branches; enabling the network to convert the detection task into a regression task; simultaneously completing the classification of the targets and the regression of the bounding boxes on one network; unifying four steps of candidate box generation, feature extraction, classification and position finishing of a target detection algorithm into a deep network framework, carrying out end-to-end training on a network model by adopting a back propagation and random gradient descent method, reducing a loss function to a small range through iterative training, and then stopping training.

Description

technical field [0001] The invention belongs to the fields of deep learning, image processing, pattern recognition, etc., and in particular relates to a deep learning-based target detection method using multiple branches to detect targets of different scales in traffic scenes. Background technique [0002] Automatically detecting various objects (such as vehicles and pedestrians) in a traffic scene is the first processing step in many intelligent transportation systems. Reasonable traffic management and control on major roads can reduce the occurrence of problems such as traffic accidents and road congestion. [0003] In the past decade, many scholars and researchers have made considerable efforts in this field and proposed some challenging benchmark datasets, such as KITTI and LSVH, etc., for evaluating and comparing various detection methods. performance of the algorithm. Since the features extracted by convolutional neural networks have better generalization performance...

AI Technical Summary

Problems solved by technology

Despite the powerful performance of CNN, one of the difficulties faced when applying CNN to object detection in traffic scenes is that traditional CNN-based methods are sensitive to scale, and the fully connected layer of CNN requires a fixed-size input, while traditional ROIPooling The scheme of simply copying some part of the proposal region to fill the extra space to obtain the feature map of the specified size will destroy the original structure of small objects.

During network training, padding replicated values ​​not only leads to inaccurate feature representation during forward propagation, but also accumulates errors during backpropagation

Inaccurate representation and accumulated errors can mislead the network training and prevent the network from correctly detecting small-scale objects

In addition, when the feature map reaches a certain depth, the small-scale object may have lost its information, which undoubtedly makes it more difficult for these methods to detect small objects accurately

Images