Unsupervised domain-adaptive scene-specific pedestrian detection for static video surveillance. (October 2021)

• Record Type: Journal Article
• Authors: Mou, Quanzheng; Wei, Longsheng; Wang, Conghao; Luo, Dapeng; He, Songze; Zhang, Jing; Xu, Huimin; Luo, Chen; Gao, Changxin
• Is Part Of: Pattern recognition. Volume 118(2021)
• Publication Date: 2021-10
• Subjects: Scene-specific pedestrian detection -- Domain adaptation -- Unsupervised learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108038
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a domain-adaptive video object detection framework to generate scene-specific pedestrian detectors in different scenarios without human-labeled target domain samples. We propose a cycle semantic transfer network to achieve adversarial domain adaptation across different surveillance videos. We propose an online gradual optimization algorithm to iteratively specialize a generic detector to a target domain. This algorithm exhibits better fault tolerance for mislabeled samples by simulating an active learning process. We conduct experiments on multiple datasets to evaluate the performance of the algorithm. The proposed self-learning method performs even better than supervised learning methods and existing scene-specific pedestrian detection methods. Abstract: Objects from one category may be drawn from different distributions due to diverse illuminations, backgrounds, and camera viewpoints. Traditional object detection methods generally perform poorly due to the domain shift. To address this problem, we propose to train a domain-adaptive scene-specific pedestrian detector in an unsupervised manner. A generic detector is transferred to different target domains from one labeled source domain dataset without human-annotated target samples. Specifically, we first extend the generic detector to a dual-boundary classifier and collect hard samples as unlabeled target samples according to the detection confidence. Then, we propose a cycle semantic transfer network to align the instance-level and class-level distributions between the source domain and target domain and automatically label the hard samples. The initial generic detector is then re-trained by these labeled hard samples and specialized to a target scene. This process can be conveniently extended to different surveillance scenarios and generate specific detectors under various static camera viewpoints. Moreover, to reduce the impact of mislabeled hard samples on the generic detector, an online gradual optimization algorithm is proposed to iteratively update the generic model, thereby obtaining an optimized process that is insensitive to individual mislabeled target samples. Extensive experiments show that even if the target domain is not manually annotated, the proposed self-learning method demonstrates the effectiveness of pedestrian detection in various domain shift scenarios, and it outperforms existing scene-specific pedestrian detection methods and some classic supervised methods.