基于三维显著性空间引导和深度网络流模型的多目标检测与小轨迹关联跟踪方法研究

Multiple object detection and tracking is one of the active discussed topics for both domestic and foreign researchers. High intensity of targets, frequent occlusion between targets, pose and scale variety always remains extremely challenging. Most of current works focus on dealing the multiple object detection and tracking problem in 2D image space, our research would intensively explore the problem of 3D multiple object detection and tracking in complex environment based on stereo vision from the perspective of human visual perception system. Inspired by human vision attention mechanism, our research proposes a new 3D visual saliency map according to the way of human visual perception mechanism, which would be applied for guiding more efficient multiple object detection in 3D space. In addition, our research proposes a deep network flow model with learning of cost function, for the task of multiple object tracking based on tracklet association, which models spatial-temporal appearance model and occlusion model, which would greatly address the track loss problem under the case of frequent occlusions. More important, regarding the cost function, which is core of the network flow, we would learn the cost function through neural network, to avoid the improper cost function which is manual designed, to further improve the tracking accuracy. This study will promote the development of theory and technology for the frontier research issues in the area of computer vision, deep learning, brain-inspired artificial intelligence.

多目标检测与跟踪是长久以来国内外学者积极研讨的富有挑战性的研究课题之一。场景内目标密集、目标间频繁遮挡、目标尺度姿态多变等情况，始终是该领域的研究难点。目前大部分工作主要集中于在二维图像层面进行多目标检测与跟踪，本课题从模拟人类视觉感知系统的思路出发，基于双目立体视觉着手解决复杂环境下三维空间多目标检测与跟踪问题。我们借鉴人类视觉注意机制建立更贴近人类大脑机制的三维视觉显著性空间，提出由三维显著性空间引导实现高效的三维多目标检测。另外，提出基于代价函数学习的深度网络流模型实现基于小轨迹关联的多目标跟踪，建模时空表观模型和目标遮挡模型，极大地解决目标物频繁被遮挡情况下目标跟踪易丢失的难题。另外，我们针对网络流中具有关键作用的代价函数采用神经网络学习的方式，避免人工设计代价函数不恰当导致跟踪结果不够理想的情况。本项目的开展有利推动计算机视觉、深度学习、类脑智能研究等领域的理论完善和技术发展。

本项目针对多目标检测与跟踪进行了深入研究，从人类视觉感知机理出发，利用双目视觉模拟人类视觉系统的运行机制，探讨了人类视觉注意中显著性特征的组织方法，提出了基于迭代式自主学习的视差估计算法，建立了贴近人类视觉注意机制的三维视觉显著性空间理论。并将显著性研究成果服务于三维目标检测任务，提出了基于特征自适应融合的三维目标检测算法。此外，进一步研究了通过端到端的方式实现基于深度关联网络模型的多目标跟踪方法，构建了基于掩膜标记的目标表观模型，为遮挡情况下目标关联提供更准确的信息。本项目从符合人类视觉感知方式的角度丰富了多目标检测与跟踪相关机制，为类脑人工智能中复杂环境感知理解的重大问题攻关提供理论基础和技术支撑。