Vehicle detection and identification method based on improved DSOD model

Abstract

The invention discloses a vehicle detection and identification method based on an improved DSOD model. The DSOD is improved on the basis of an SSD algorithm, can be simply understood as SSD + DenseNet= DSOD, adopts proposal-free detection model SSD, and adds a DenseNet thought. The DSOD model is divided into two parts, namely a Backbone for feature extraction and a Front-end for target prediction. The Backbone sub-network is similar to a DenseNet, is composed of a layer of Stem block (main module), four layers of Dense blocks (Dense modules), two layers of Translation layers (transition layers) and two layers of Translation w / o pooling layers (transition w / o pooling layers), and is used for extracting image features. The Fort_end sub-network_network (front-end detection sub-network) implements a border frame detection effect through Dense Connetion.

Description

technical field [0001] The invention relates to the technical field of traffic image processing, in particular to a vehicle detection and recognition method based on an improved DSOD model. Background technique [0002] With the development of the national economy and the development of artificial intelligence technology, the field of intelligent driving vehicles has achieved unprecedented development. Target detection is a key part of the environment recognition and perception of intelligent driving vehicles. The detection of target vehicles ahead of the road helps to make safe and accurate driving decision-making instructions. In the target detection task of computer vision in daily driving scenarios, the road conditions are quite complex, and unfavorable factors such as vehicles of various sizes, mutual occlusion of vehicles, and changes in light in sunny and rainy days all affect the accuracy of vehicle target detection. In the field of target detection, the cost of mak...

AI Technical Summary

Problems solved by technology

The Faster R-CNN method belongs to two-level detection. It uses Region Proposal Networks (RPN) to extract a large number of candidate regions, classify and regress them, and obtain the location of the target. However, due to the slow operation speed and memory Huge consumption

The YOLO (You only look once, YOLO) method belongs to the first-level detection. It uses a 7*7 network for end-to-end regression calculation, predicts the frame coordinates in each grid and the confidence of each category, and combines the classification and positioning of the target. , although the detection speed index is relatively fast, the detection accuracy of the simple network regression operation is not high

SSD (Single Shot MultiBox Detector, SSD) is also an end-to-end algorithm that introduces different convolutional layer detection structure predictions to improve multi-scale and multi-scale detection capabilities, but the lack of low-level feature semantic information leads to weak small target detection capabilities

Images