监控场景中的实用目标检测方法研究

Object detection in surveillance (including pedestrian and vehicle detection) is a fundamental problem in intelligent video analysis, whose performance is directly related to the success of the high-level video analysis. Although object detection has achieved great success in specific scenario, it is still a challenging problem due to the variations of occlusion, low-resolution, large appearance deformation etc. This project proposes to improve the object detection performance in unconstrained complex scenes. In particular, we propose to exploit the image/video contextual information to improve the detection performance in occlusion. A common discriminant feature extraction method is proposed to enhance the performance with different resolutions. An adaptive object detection mechanism is developed to improve the utility of object detection in different scenarios. This project also investigates the fast object detection algorithm to make the object detection effective and efficient in practice.

监控场景中的目标检测（主要包括行人和车辆检测）是智能视频分析中的核心基础问题。目标检测性能的好坏直接关系着上层智能视频分析的性能。现有的目标检测方法在特定场景下已经取得了较好的检测性能，却也仍旧存在一系列待解决的难点问题。目标检测性能受低分辨率，遮挡，场景变化等因素影响较大，检测速度较慢等均是监控场景下的现有技术仍未完全解决的问题。本课题针对上述难点问题展开研究，通过研究图像和视频上下文信息，提高遮挡情况下的目标检测率。通过共同判别特征学习，提高不同分辨率、不同表象形变下的目标检测性能。通过研究场景自适应算法，提高目标检测算法的通用性和便捷性。此外，本课题还进一步研究探索快速目标检测算法，在不明显影响检测精度的前提下，降低目标检测算法复杂度，提升算法的实用性。

本项目针对监控场景中目标检测的多分辨率，小尺度目标，场景自适应，目标遮挡，大姿态关键点定位等难点问题展开研究，研究开发了基于深度学习的端到端目标检测技术，包括1）GPU实时的高精度多分辨率人脸检测器；2）CPU实时的高性能人脸检测器；3）遮挡鲁棒的行人检测方法；4）场景自适应的目标检测方法；5）大姿态下的人脸关键点定位算法等。相关技术在国际公开测试集上均取得了同时期国际领先的结果，相关工作整理发表在国际顶级期刊和顶级会议上。本项目共发表论文24篇，申请发明专利4项，项目成果有效推动了监控场景下目标检测技术的发展，具有较好的实用价值。