A NOVEL DEPTHWISE CONVOLUTIONAL VARIATIONAL AUTOENCODER FOR ANOMALY DETECTION IN COMPLEX TRAFFIC SCENARIOS FROM UAV VIEWS

Date

2025-8-13

Author

Saleem, Arslan

Metadata

Show full item record

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

Item Usage Stats

82
views

0
downloads

Cite This

Traffic anomaly detection (AD) is essential for improving public safety, reducing risks, and enabling quick responses in intelligent surveillance systems. Aerial traffic monitoring, particularly using drones, has gained attention due to its potential to address challenges like dynamic urban environments, yet it remains underexplored. Detecting anomalies in drone captured video involves unique obstacles: rare events, small and overlapping objects, multi-scale targets, and complex backgrounds. To address these challenges, we propose the Depthwise Convolutional Variational Autoencoder (DwCVAE), a novel model designed to enhance AD in drone-based traffic surveillance. DwCVAE leverages depthwise convolutions, which allow efficient and detailed feature extraction, improving model sensitivity to subtle and multi-scale anomalies. Built on variational autoencoder (VAE) architecture, DwCVAE creates compact latent representations that capture normal traffic patterns, enabling reliable detection of deviations. This depthwise approach marks a key innovation, optimizing both computational efficiency and detection accuracy. We design four additional models: Convolutional Variational Autoencoder (CVAE), Dilated Convolutional VAE (DCVAE), Separable Convolutional VAE (SCVAE), and Convolutional LSTM VAE (CLSTMVAE) to systematically assess the effectiveness of DwCVAE. Additionally, we evaluate DwCVAE against state-of-the-art weakly supervised and unsupervised models on two benchmark datasets, Drone-Anomaly and UIT-Adrone. Experimental results show that DwCVAE outperforms competing models, achieving an AUC of 74.95 and 79.77 and an EER of 0.30 and 0.27 on Drone-Anomaly and UITAdrone, respectively, demonstrating its superior performance in complex aerial surveillance tasks.

Subject Keywords

Anomaly Detection, Depthwise Convolutional, Traffic Surveillance, Unsupervised Learning, Variational Autoencoder

URI

https://hdl.handle.net/11511/116183

Collections

Northern Cyprus Campus, Thesis