BirdView Retina-Net: Small-Scale Object Detector for Unmanned Aerial Vehicles

Abstract

Recent advancements in computer vision and deep learning techniques have led to many applications in robotics. These techniques play a significant part in the development of artificially intelligent autonomous and semi-autonomous robotic systems, e.g., Unmanned Aerial Vehicles (UAVs) and autonomous cars. Object detection and tracking is an important pillar for such robotic systems. For object detection, a wide variety of deep learning frameworks have been proposed over the past few years. However, objects in the video from a UAV are small and thus very hard to localize and classify. This research proposes a deep learning framework based object detection approach named as BirdView Retina-Net (BV-RNet). BV-RNet is an object detection framework capable of efficiently detecting small-scaled objects from an aerial view-point. BV-RNet extracts dense features and optimizes pre-defined anchors for inference to learn features effectively. VisDrone dataset is used to evaluate the performance of the framework. We have compared the results of BV-RNet with past VisDrone challenge competitors and other commonly employed past object detection frameworks. We have reported that our algorithm achieves mean average precision (mAP) of 31.8 on test-dev dataset, which is among the highest mAP achieved so far on VisDrone. Furthermore, it also outperforms several of the top class-wise mAP of various classes when compared with VisDrone past challenge competitors.