Small Object Detection Without Attention for Aerial Surveillance

Abstract

This paper introduces the development of an essential deep-learning model for surveillance systems utilizing high-mounted CCTV or drones. Objects seen from elevated angles often look smaller and may appear at different angles compared to ground-level observations. To improve the detection of small objects, we propose a network incorporating an element-wise multiplication module based on the vanilla Vision Transformer (ViT) architecture. However, traditional transformer models need significant computational resources, which may not be practical for edge devices like CCTV cameras or drones. Therefore, we apply the Attention-Free Transformer (AFT) to reduce computational requirements enabling real-time operation on low-capacity devices. We validate the performance by combining ViT and AFT with the YOLOv5 real-time object detection model. Practical applicability is confirmed by implementing it on the low-capacity device named ODROID H3+. Validation datasets include Autonomous Driving Drone, VisDrone, AerialMaritime, and PKLot, all containing numerous small-sized objects. Experimental results on the VisDrone dataset show that YOLOv5 nano + AFT reduces parameter count by 4.6% while increasing accuracy by 1%, making it an efficient network. The model size is suitable for edge device implementation at 3.7 MB. Similarly, Aerial Maritime and PKLot datasets indicate a decreased amount of parameters and increased accuracy. Hence, the proposed deep learning model is applicable for aerial surveillance systems.