Road Object
Classification

What We Did

Our project followed a full road-object image classification pipeline, which we later repeated after improving the realism and quality of the dataset. We began by gathering labeled object images and organizing them by class. Each image was resized to a common resolution and preprocessed through steps such as grayscale conversion, normalization, and data cleaning to make the inputs more consistent. We then compared multiple feature representations, including raw pixel values, edge-based features, frequency-domain features derived from the discrete Fourier transform, and color-based features. Using these representations, we trained and evaluated four MATLAB classifiers: KNN, Naive Bayes, SVM, and Decision Tree. Model performance was measured using confusion matrices, accuracy comparisons, and analysis across both balanced and refined real-world datasets. After finding that the original dataset contained unrealistic and suboptimal images, we replaced many of those examples with more representative dashcam-style objects and reran the classification pipeline to observe how dataset realism changed the results.

What We
Classified

Initially, we struggled finidng good datasets with cropped dash-cam images of the classes we desired. So we used some suboptimal datasets while we procured and produced our own dataset.
We used labeled datasets from Kaggle and dashcam-related sources. Initial experiments balanced all classes at 50 images each. Later runs used unbalanced real-world proportions (e.g. 414 cars, 234 pedestrians, 3029 signs) to better reflect deployment conditions.

To improve realism, we moved away from an older car dataset after finding many images that were not representative of dashcam footage. It contained luxury vehicles (e.g., Rolls-Royces), heavily edited promotional images, text overlays, and staged settings. Using such images risks teaching the model non-generalizable patterns that would hurt roadside detection in practice. BadCar(43).jpg is an example of an image that triggered this refinement.

We created part of our own dataset by cropping objects from authentic dashcam footage, improving perspective, lighting, motion blur, and framing consistency. However, this custom set has limitations: many clips were sourced from Russia and fewer from the United States, Canada, and China. Pedestrian data may underrepresent demographic, clothing, and environment diversity in North America, which could produce uneven model performance across groups and regions. pedestrian46.jpg exemplifies this limitation. Expanding to broader demographic and geographic data is an important next step.

All images were in JPG format for their smaller file size and prevalence in real dashcam footage.

• Bicycle — Kaggle balanced subset
• Motorcycle — Kaggle balanced subset
• Vehicle / Car — 414 images (unbalanced run)
• Pedestrian — 234 images (unbalanced run)
• Traffic Sign — 3,029 images (unbalanced run)
• Traffic Light — Kaggle balanced subset

Car · Pedestrian · Sign

With a larger, refined, and unbalanced dataset that accurately reflects realworld data. Most classifiers perform dramatically better — especially SVM, which achieves over 91% accuracy. However, Naive Bayes performed significantly worse than any other classifier which revealed a fundamental flaw. That the Naive Bayes classifier will assign a prior probability based on how often a class appears in training sets.