A CASCADED PART-BASED SYSTEM FOR FINE GRAINED VEHICLE CLASSIFICATION

Abstract

Vehicle make and model recognition (VMMR) has become an important part of intelligent transportation systems. VMMR can be useful when license plate recognition is not feasible or fake number plates are used. VMMR is a hard, finegrained classification problem, due to the large number of classes, substantial inner-class, and small inter-class distance. A novel cascaded part-based system has been proposed in this paper for VMMR. This system uses latent support vector machine formulation for automatically finding the discriminative parts of each vehicle category. At the same time, it learns a partbased model for each category. Our approach employs a new training procedure, a novel greedy parts localization, and a practical multi-class data mining algorithm. In order to speed up the system processing time, a novel cascading scheme has been proposed. This cascading scheme applies classifiers to the input image in a sequential manner, based on the two proposed criteria: confidence and frequency. The cascaded system can run up to 80% faster with analogous accuracy in comparison with the noncascaded system. The extensive experiments on our data set and the CompCars data set indicate the outstanding performance of our approach. The proposed approach achieves an average accuracy of 97.01% on our challenging data set and an average accuracy of 95.55% on CompCars data set.

Existing System

Holistic Approaches, Part-Based Approaches

Proposed System

The proposed training procedure employs a combination of standard support vector machine (SVM) and Latent SVM algorithms to learn a model per class of vehicles, and finds the discriminative parts of each class simultaneously. Each model contains a root filter, part filters and the spatial relationship of parts. Root filter captures global appearances, and part filters capture local appearance properties of a specific vehicle class. The spatial relationship between parts helps a model to distinguish between similar classes with a different configuration of parts. Such a model can be learned using Latent SVM or MI-SVM algorithms. These individual models are used together with a max-voting scheme. However, the models aren’t calibrated and their scores can’t be compared to each other as they are. Thus, we used Platt-Scaling and fitted a logistic regression model to all of the classifiers scores. Furthermore, in order to decrease the system processing time, a novel cascading scheme has been proposed. This cascading scheme can control the trade-off between accuracy and speed of the system, by applying the classifiers in a sequential manner to the input. We proposed a practical ordering of classifiers based on two criteria, namely confidence, and frequency. For further boosting the system speed, each model in the cascaded system is replaced with its cascaded version. With this modification, the system processing time is reduced by 93%.

Architecture

BLOCK DIAGRAM