Small vehicles detection based on UAV

Abstract

In recent years, with the development of deep neural networks, object detection has been widely used in many fields such as video surveillance, face recognition, and unmanned aerial vehicles (UAVs). However, small objects detection is still a challenge for the most of existing methods due to its low resolution and limited information. Here, the authors use an effective object detector single shot multi-box detector (SSD) as base architecture to detect small vehicles shot by UAV platform. First, in order to improve the performance for small objects, an extended-layer feature method is proposed for introducing more small-scale information. Second, the shapes of default boxes on specific feature maps are adjusted to balance the performance loss caused by adding feature maps. Experimental results show that this method reached the accuracy of 78.6% on PASCAL VOC object-detection dataset, which has a 1.8% bonus compared with the baseline. Moreover, in order to show the effectiveness on small vehicles, an additional experiment is conducted on the UAV123 dataset, which gets a 3.1% bonus. The speed of the proposed method has a 39 fps on a NVIDIA GTX Titan X.

Keywords

• object detection
• feature extraction
• neural nets
• autonomous aerial vehicles
• vehicles detection
• deep neural networks
• object detection
• video surveillance
• face recognition
• unmanned aerial vehicles
• effective object detector single shot multibox detector
• base architecture
• uav platform
• extended-layer feature method
• small-scale information
• default boxes
• pascal voc object-detection dataset