PEDESTRIAN PRESENCE DETECTION ON A CROSSWALK USING A FIXED ROADSIDE CAMERA: A DEPLOYMENT ORIENTED EVALUATION OF CLASSICAL AND DEEP LEARNING APPROACHES

Date

2026-7-7

Author

Haq, Muhammad Arslan-Ul-

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

108
views

0
downloads

Cite This

Pedestrian safety at urban crosswalks remains a key concern in intelligent transport systems, particularly for surveillance-based applications that require both accurate detection and real-world efficiency. This dissertation investigates crosswalk pedestrian presence detection using a stationary roadside camera and a region-of-interest based binary classification approach. The aim is to determine whether at least one pedestrian is present in the crosswalk region of a given frame, while maintaining strong accuracy and CPU-based computational efficiency. A deployment-oriented evaluation framework was developed using a selected fixed roadside sequence from the StreetAware dataset. Frames were processed through region-of-interest masking and cropping. A task-specific labeling strategy was applied based on whether the bottom center point of a pedestrian ground truth bounding box lay inside the predefined crosswalk region. Traditional machine learning and deep learning methods were evaluated under the same dataset split, preprocessing pipeline, and CPU-based testing protocol. The traditional branch included HOG, LBP, combined features, and several variants with SVM, Random Forest, Naive Bayes, KNN, and XGBoost. The deep learning branch included ResNet18 with and without pretrained weights, and a proposed lightweight custom CNN. The pretrained ResNet18 achieved 0.95 accuracy, 0.95 recall, 0.97 F1-score, and 0.99 Average Precision. The proposed CNN achieved 0.91 accuracy, 0.90 recall, 0.92 F1-score, 0.98 Average Precision, and 21.18 FPS on CPU. The best handcrafted model, XGBoost with HOG and LBP, achieved 0.90 F1-score, 0.97 Average Precision, and 22.61 FPS. Overall, pretrained ResNet18 provided the strongest predictive reliability, while the proposed CNN offered the best balance between accuracy and efficiency.

Subject Keywords

Deployment oriented evaluation, Fixed roadside camera, Lightweight CNN, Pedestrian presence detection on a crosswalk, ROI based classification

URI

https://hdl.handle.net/11511/120088

Collections

Northern Cyprus Campus, Thesis