Far-infrared pedestrians detection based on adaptive template matching and heterogeneous-feature-based classification
Guohua Wang, Qiong Liu, Yongsen Zheng, Shaowu Peng · 2016
Robust and efficient far-infrared pedestrian detection is challenging for advanced driver-assistance systems (ADAS). In order to address problems that the performance of existing detection methods often suffers from uncontrolled background objects, various pedestrian size and pose, a novel pedestrian detection method that integrates the improved template-matching and machine-learning techniques is proposed to address the above two problems. With the purpose of reducing the interference of uncontrolled background objects, we employ template matching technique to pre-classify candidates so that false alarms can be reduced. During this procedure, we create more templates and design a head location algorithm to adaptively select an optimal head template. To reduce the interference of various pedestrian size and pose, we propose a novel heterogeneous feature which concatenates several state-of-the-art features to enhance the pedestrian description ability, then the heterogeneous feature is fed to a machine learning classifier to validate pedestrians from the candidates. Experiments in public and additional datasets captured by ourselves indicate that, compared with three state-of-the-art methods, the proposed method can obtain higher detection accuracy.