Two-stream small-scale pedestrian detection network with feature aggregation for drone-view videos

• Published in: Multidimensional systems and signal processing Vol. 32; no. 3; pp. 897 - 913
• Main Authors: Xie, Han, Shin, Hyunchul
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2021; Springer Nature B.V
• Subjects: Artificial Intelligence; Circuits and Systems; Computer networks; Datasets; Drone aircraft; Electrical Engineering; Engineering; Feature maps; Flow nets; Frame design; Pedestrians; Signal,Image and Speech Processing; Deep learning; Pedestrian detection; Neural network; Feature aggregation; Drone vision
• ISSN: 0923-6082; 1573-0824
• DOI: 10.1007/s11045-021-00764-1

Detecting small-scale pedestrians in aerial images is a challenging task that can be difficult even for humans. Observing that the single image based method cannot achieve robust performance because of the poor visual cues of small instances. Considering that multiple frames may provide more information to detect such difficult case instead of only single frame, we design a novel video based pedestrian detection method with a two-stream network pipeline to fully utilize the temporal and contextual information of a video. An aggregated feature map is proposed to absorb the spatial and temporal information with the help of spatial and temporal sub-networks. To better capture motion information, a more refined flow net (SPyNet) is adopted instead of a simple flownet. In the spatial stream subnetwork, we modified the backbone network structure by increasing the feature map resolution with relatively larger receptive field to make it suitable for small-scale detection. Experimental results based on drone video datasets demonstrate that our approach improves detection accuracy in the case of small-scale instances and reduces false positive detections. By exploiting the temporal information and aggregating the feature maps, our two-stream method improves the detection performance by 8.48% in mean Average Precision (mAP) from that of the basic single stream R-FCN method, and it outperforms the state-of-the-art method by 3.09% on the Okutama Human-action dataset.