SPID: Surveillance Pedestrian Image Dataset and Performance Evaluation for Pedestrian Detection

Pedestrian detection is highly valued in intelligent surveillance systems. Most existing pedestrian datasets are autonomously collected from non-surveillance videos, which result in significant data differences between the self-collected data and practical surveillance data. The data differences include: resolution, illumination, view point, and occlusion. Due to the data differences, most existing pedestrian detection algorithms based on traditional datasets can hardly be adopted to surveillance applications directly. To fill the gap, one surveillance pedestrian image dataset (SPID), in which all the images were collected from the on-using surveillance systems, was constructed and used to evaluate the existing pedestrian detection (PD) methods. The dataset covers various surveillance scenes and pedestrian scales, view points, and illuminations. Four traditional PD algorithms using hand-crafted features and one deep-learning-model based deep PD methods are adopted to evaluate their performance on the SPID and some well-known existing pedestrian datasets, such as INRIA and Caltech. The experimental ROC curves show that: The performance of all these algorithms tested on SPID is worse than that on INRIA dataset and Caltech dataset, which also proves that the data differences between non-surveillance data and real surveillance data will induce the decreasing of PD performance. The main factors include scale, view point, illumination and occlusion. Thus the specific surveillance pedestrian dataset is very necessary. We believe that the release of SPID can stimulate innovative research on the challenging and important surveillance pedestrian detection problem. SPID is available online at: http://ivlab.sjtu.edu.cn/best/Data/List/Datasets.