Vehicle sticker recognition based on multi-feature encoding and feature matrix distance

Abstract

A novel algorithm to use vehicle sticker (or tag) features and encode the features is proposed. It can make the representation more precise and recognition more accurate. In vehicle recognition or searching, traditional algorithms will be limited because they focus only on the features extracted from colors, logos or sub-types that are not enough to identify a vehicle. Furthermore, the license plate (LP) can be forged easily so the LP is not reliable to identify a specified vehicle. Our algorithm solves this problem by sticker multi-feature encoding. Most vehicles have printed permission labels or certification symbols named vehicle stickers or tags mounted on the frontal glass. These stickers are a kind of special fingerprint features to identify a unique vehicle. Every driver has his own habit to paste different stickers. In this meaning these stickers form specified multi-feature including color, shape, position and amount. Our algorithm encodes the sticker multi-feature to construct structured feature presentation, i.e. the sticker code. In recognition stage, with the matrix distance of the multi-feature encoding, the detailed sticker code can be utilized to distinguish the vehicle types and colors reliably, and can recognize the tiny difference among vehicles with the same colors, logos and even sub-types. Our algorithm decreases the amount of vehicle candidates effectively by accurate feature coding. In our experiments, we coped with 10000 vehicle images taken by public traffic surveillance system to verify the effectiveness of this algorithm in vehicle sticker multi-feature encoding recognition.