基于显著性混合级联和深度网络树的智能驾驶多目标检测与跟踪

In driving environment, pedestrians, bicycles, motorcycles, electric motorcycles, and other non-four-wheel-vehicle targets are the biggest victims in urban traffic accidents, and are the most important goals in the detection and avoiding of intelligent driving systems. Focusing on solving the problem of lacking generalization ability of traditional methods in multi-target detection, this project will study the multi-target detection of non-four-wheel-vehicle targets; taking full account of the specificity of the vehicle driving systems, we will study the integration method of multi-objective detection and tracking. There are main three parts as following. Firstly, this study aims to construct a visual saliency model according to the driver's cognitive characteristics; with learned visual saliency features, we will study an asymmetric Boosting learning method to train saliency hybrid cascade structure to achieve region of interest extraction. Secondly, the construction and training method of convolution kernel of deep convolution network are to be studied to reduce the convolution network redundancy; according to the similarity degree of different targets, we aim to construct a cascade tree structure of convolution network to realize multi-objective detection in a coarse-to-fine process. Thirdly, a method of association feedback tracking is to be studied, which constructs a binary potential function to correlate and track the detection results; and a feedback of the tracking results is designed to direct the detection process to realize the integration of detection and tracking. The algorithm is designed to solve the multi-target detection and tracking problem in intelligent driving scene, and to provide theoretical and technical support for intelligent driving system research.

行车环境中的行人、自行车、摩托车和两轮电动车等非四轮车目标是城市交通事故的最大受害者，也是智能驾驶系统最需要检测和避让的目标之一。本课题研究非四轮车目标检测方法，重点解决传统检测方法在多目标检测中泛化能力不足的问题，并充分考虑车载系统的特殊性，研究多目标的检测跟踪一体化方法。主要包括：构建符合驾驶员认知特点的视觉显著特征，并针对非均衡随机负样本研究一种非对称Boosting学习方法，来训练显著特征混合级联结构，实现感兴趣区域提取；研究深度卷积网络的卷积特征挖掘方法以减少卷积网络冗余，并根据目标相似度构建卷积网络的级联树型结构，实现多目标的由粗到精快速检测；研究一种关联反馈跟踪方法，构造二元势函数将检测结果进行数据关联和跟踪，并将跟踪结果进行反馈融合，实现检测跟踪一体化。本项目所研究的算法旨在解决智能驾驶场景中的多目标检测跟踪问题，为智能驾驶系统研究提供理论支撑和技术保障。

行车环境中的行人、自行车、摩托车和两轮电动车等非四轮车目标是城市交通事故的最大受害者，也是智能驾驶系统最需要检测和避让的目标之一。本课题研究非四轮车目标实时检测跟踪方法，重点解决传统检测方法在多目标检测中泛化能力不足和速度慢的问题，并充分考虑车载系统的特殊性，研究多目标的检测跟踪一体化方法。主要研究内容包括：.1）受人类视觉显著性快速高效机制启发，研究一种贝叶斯树级联ACF树形结构，得到驾驶场景中符合驾驶员视觉认知的显著性特征和优先次序。针对训练样本的高度非均衡性和随机性特点，研究一种面向非均衡随机负样本的非对称Boosting训练方法，使正样本与负样本采用不同误差模型进行训练和收敛，以减少数目巨大的随机性负样本对训练结果的负面影响。实验结果表明：该方法感兴趣提取率达到100%，且训练方法可适用于未知小样本。.2）为解决特征冗余、节点数量大、平台要求高的问题，研究一种基于PY-YOLO网络混合级联模型，将检测目标按其结构和尺寸进行混合级联分配，以达到相似目标卷积核共享和进一步消除冗余的目的。对树结构的每一个节点分别训练YOLO网络，使整个级联卷积网络能够进行由粗到精检测，快速排除部分背景样本并保留目标样本，提高算法整体的表达能力和计算速度。实验结果表明：该方法的平均精确度AP值比YOLO等方法高13%。.3）研究一种数据关联互反馈跟踪方法，根据检测过程中的分块方法对目标进行分块，并将检测目标的分块卷积核作为跟踪特征，通过构造二元势函数对主分块特征的相容性概率和互斥性概率进行建模，实现多目标的数据关联匹配和跟踪，既减少了跟踪特征时间又提高了鲁棒性。实验结果表明：该方法在FAF、MOTP和MOTA等多个指标均比其他对比方法有更好表现。.4）采用主辅双目视觉系统实现运动目标的快速检测、定位和跟踪，并在车载辅助驾驶系统的行车环境中进行了实验验证，最终形成一套适合多目标检测跟踪及预警的辅助驾驶系统。