Compressive Detection for Camera Array Images

Abstract

High-resolution camera arrays have been used in wide area and long distance surveillance as well as event recording. However, processing and storing the huge video streams of camera arrays remain a heavy burden for many applications. This paper presents a compressive object detection framework to reduce the camera data volume, and to accelerate the detection using high-resolution images with small performance degradation. The proposed method superimposes multiple images from different sub-cameras of the array, and performs detection on the superimposed data using neural networks. Detected bounding boxes are then relocated on the original sub-images as candidates which are further verified through target classification. The system only stores the high-resolution superimposed data and the low-resolution wide FOV images, which can guarantee the detection accuracy with smaller data volume. The proposed methods are validated using pedestrian datasets and real camera array images in terms of detection accuracy.