An Improved CornerNet-Lite Method for Pedestrian Detection of Unmanned Aerial Vehicle Images

Pedestrian detection is a key component to analyze the outdoor scene, which has an important part in object detection of unmanned aerial vehicle (UAV) images. The original ComerNet-Lite is sensitive to shooting angle and environments, making it difficult to detect pedestrians in case of distortion, blur and small scale. To address this issue, this paper proposes an improved ComerNet-Lite framework, which can effectively detect pedestrians in unknown angles and complex environments. First, CornerNet-Lite is improved by introducing the stacked hourglass module into the backbone to raise the ability to capture features and extracting information at every scale of UAV images. Second, we employ soft non maximum suppression (Soft-NMS) to improve pedestrian detection accuracy by modifying the confidence and detection box of the adjacent pedestrian. Finally, we evaluate improved CornerNet-Lite on the VisDrone2019 dataset achieves satisfactory results.