Dense and Small Object Detection in UAV Vision Based on Cascade Network

Abstract

With the development of Unmanned Aerial Vehicles, drones are being deployed in a number of commercial and civil government applications ranging from remote surveillance and infrastructure maintenance among others. However, processing the videos captured by drones for the extracting meaningful information is hindered by multitude of challenges that include, the appearance of small objects, changes in viewpoint of these objects, illumination changes, large-scale resolution of the captured video, occlusion and truncation. Addressing these challenges, there is a critical need to develop algorithms that is able to efficiently process the videos that can result in robust detection and recognition of small objects. In this paper, we propose a novel processing pipeline, that brings together several key contributions including (i) the introduction of DeForm convolution layers within backbone; (ii) use of the interleaved cascade architecture; (iii) data augmentation process based on crop functionality and (iv) multi-model fusion of sub-category detection networks. The proposed approach has been exhaustively benchmarked against VisDrone-DET object detection dataset, which includes 10,209 images for training, validation and testing. The evaluation of the proposed approach has resulted in 22.61 average precision on the test-challenge set in VisDrone-DET 2019.