Strategies for Improved Out-of-Distribution Detection in Drone vs. Bird Classification

The use of drones has expanded significantly across various applications over the past decade, leading to increased surveillance-related challenges. These challenges raised the necessity of developing Anti-Drone systems. One of the critical requirements for an effective Anti-Drone system is the ability to accurately distinguish drones from birds in the sky. While deep learning-based classification techniques have been employed for this task, they often suffer from high false positive rates. To address this challenge, Out-of-Distribution (OOD) detection is essential for enhancing the reliability and robustness of drone surveillance systems, particularly in differentiating drones from birds. This paper explores several techniques to improve OOD detection performance, focusing on Energy-Based Models (EBM) and Variational Autoencoders (VAE). We evaluate four loss functions within the EBM framework: Mean Squared Error (MSE) Loss, Mean Squared Error with OOD Penalty, Contrastive Loss, and Binary Cross-Entropy with Energy Regularization. Our results demonstrate that the Mean Squared Error with OOD Penalty function achieves the best performance, with an AUC of 0.9, providing clearer separation between in-distribution (drones) and out-of-distribution (birds) samples. However, the VAE approach did not yield significant results for the binary classification task. Future work could explore hybrid approaches to further enhance OOD detection in such applications.