Vehicle type recognition method based on deep Fisher network

Abstract

The invention provides a vehicle type recognition method based on a deep Fisher network. Firstly the 0th layer of the Fisher network is constructed, and the SIFT features of each type of vehicle type image are extracted as for a database having K types of vehicle type images; then the 1st layer of the Fisher network is constructed, Fisher vector coding is performed on the extracted SIFT features, the coded vectors are spatially stacked and then L2 normalization and PCA dimension reduction are performed; Fisher vector coding is performed on the obtained feature expressions of the 1st layer, and the 2nd layer of the Fisher network is formed by symbolic square root and L2 normalization processing; finally a global feature expression obtained through different vehicle type images is applied to linear support vector machine training so that a recognition system having K vehicle types is obtained; and as for a vehicle to be recognized, the vehicle to be recognized is enabled to pass the Fisher network to obtain test feature vectors, and the test feature vectors are imported to the recognition system so that the type of the vehicle to be recognized can be recognized.

Description

technical field [0001] The invention belongs to the technical field of pattern classification, and in particular relates to a vehicle identification method based on a deep Fisher network. Background of the invention [0002] Traffic information collection is the basis for building a dynamic traffic information platform for intelligent transportation systems, and vehicle types are an important part of traffic information. Road and bridge, parking lot charging system, road and bridge management and monitoring system, etc. all need vehicle type identification. In the intelligent traffic management system, the vehicle type recognition system can automatically and real-time detect passing vehicles and identify the vehicle type, license plate, and vehicle logo of the traffic management system, which can be widely used in road vehicle information records, expressway automatic toll collection, electronic police Monitoring, parking lot safety management, accident, suspicion, trackin...

AI Technical Summary

Problems solved by technology

However, these methods are difficult to solve the contradiction between complexity and generalization in pattern recognition. Although neural networks have powerful modeling capabilities, they classify large-scale images, and their huge parameter space makes it difficult to find excellent optimal initial values.

Images