Home / Current Issue / Paper 1717084
Efficient Net-B3 with Convolutional Block Attention and Focal Loss for Road Accident Detection in CCTV Surveillance Videos
Subject area: Science,Engineering and Technology · Area of research: Computer Vision, Deep Learning
DOI: https://doi.org/10.64388/IREV9I10-1717084
Abstract
Automated road accident detection from CCTV footage is a critical public safety challenge in smart city environments. Prior work using finetuned AlexNet on the ckay16 dataset achieved only 68.0% accuracy, a true positive rate (TPR) of 77.4%, and a critically high false positive rate (FPR) of 42.6% — rendering the system impractical for live deployment. In this paper we present a fully algorithmic framework that addresses these shortcomings without collecting additional data. Our method replaces AlexNet with an EfficientNet-B3 backbone, attaches a Convolutional Block Attention Module (CBAM) to the final feature block, and trains using Focal Loss with label smoothing, a Weighted Random Sampler, and a cosine-annealing learning-rate schedule with warm restarts. At inference, five-crop Test-Time Augmentation (TTA) further reduces false positives. On the held-out ckay16 test set the system achieves 96.0% accuracy, 95.7% TPR, 3.8% FPR, macro-F1 of 96.0%, and ROC-AUC of 0.981 — a 28-point accuracy gain and 39-point FPR reduction over the AlexNet baseline, demonstrating significant improvement over existing baselines.
Keywords
Road accident detection, EfficientNet, CBAM, Attention mechanism, Focal loss, Test-time augmentation, Transfer learning, CCTV surveillance
How to cite this paper
@article{1717084,
author = {M Koteswara Rao, G Raja Sekhar Reddy, P Swarna Kamal, P S C S V Sainadh},
title = {Efficient Net-B3 with Convolutional Block Attention and Focal Loss for Road Accident Detection in CCTV Surveillance Videos},
journal = {Iconic Research And Engineering Journals},
year = {2026},
volume = {9},
number = {10},
pages = {3725-3731},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1717084.pdf},
abstract = {Automated road accident detection from CCTV footage is a critical public safety challenge in smart city environments. Prior work using finetuned AlexNet on the ckay16 dataset achieved only 68.0% accuracy, a true positive rate (TPR) of 77.4%, and a critically high false positive rate (FPR) of 42.6% — rendering the system impractical for live deployment. In this paper we present a fully algorithmic framework that addresses these shortcomings without collecting additional data. Our method replaces AlexNet with an EfficientNet-B3 backbone, attaches a Convolutional Block Attention Module (CBAM) to the final feature block, and trains using Focal Loss with label smoothing, a Weighted Random Sampler, and a cosine-annealing learning-rate schedule with warm restarts. At inference, five-crop Test-Time Augmentation (TTA) further reduces false positives. On the held-out ckay16 test set the system achieves 96.0% accuracy, 95.7% TPR, 3.8% FPR, macro-F1 of 96.0%, and ROC-AUC of 0.981 — a 28-point accuracy gain and 39-point FPR reduction over the AlexNet baseline, demonstrating significant improvement over existing baselines.},
keywords = {Road accident detection, EfficientNet, CBAM, Attention mechanism, Focal loss, Test-time augmentation, Transfer learning, CCTV surveillance},
month = {April},
doi = {https://doi.org/10.64388/IREV9I10-1717084}
}