Vehicle Make and Model Recognition using Bag of Expressions

Area of Science:

• Computer Vision
• Machine Learning
• Intelligent Transport Systems

Background:

• Vehicle make and model recognition (VMMR) is crucial for automated vehicular surveillance (AVS) and intelligent transport systems (ITS).
• Existing methods like Bag of Words (BoW) have limitations in handling real-world variations.
• There is a need for robust VMMR techniques that can manage occlusion, scale, and view variations.

Purpose of the Study:

• To propose and evaluate the suitability of the Bag of Expressions (BoE) approach for VMMR applications.
• To enhance VMMR performance by integrating neighborhood information with visual words.
• To demonstrate the effectiveness of BoE in improving occlusion handling, scale invariance, and view independence.

Main Methods:

• Feature extraction using a combination of keypoint detectors and Histogram of Oriented Gradients (HOG) descriptor.
• Formation of an optimized dictionary of expressions via k-means clustering of visual words.
• Classification using multiclass linear Support Vector Machines (SVM) trained on BoE features.

Main Results:

• The BoE approach demonstrated superior performance compared to recent VMMR methods on the NTOU-MMR dataset.
• Achieved promising average accuracy and processing speed, indicating real-time applicability.
• The integration of neighborhood information significantly improved recognition capabilities.

Conclusions:

• The Bag of Expressions (BoE) approach is highly suitable for vehicle make and model recognition (VMMR) tasks.
• BoE offers significant advantages over traditional methods, particularly in challenging conditions.
• The proposed method is a viable solution for real-time VMMR systems in intelligent transport and surveillance.