基于Stixel-world及特征融合的双目立体视觉行人检测

发布时间：2018-05-18 08:05

本文选题：行人检测 + 双目立体视觉　；参考：《仪器仪表学报》2017年11期

【摘要】：针对单目视觉行人检测无法获得深度信息从而导致冗余信息较多、检测效率和准确度存在局限性的问题,首先,在图像的预处理阶段提出了一种利用双目立体视觉产生的视差信息优化分析来简化复杂场景的动态规划棒状像素场景(stixel-world)表达方式;然后,在行人目标检测阶段,对传统HOG特征中block尺度进行分析、降维,采用Fisher准则筛选得到了适用于道路环境下的多尺度HOG(multi-HOG)特征,将Multi-HOG特征与LUV颜色通道特征进行融合,最后采用交叉核支持向量机(hikSVM)分类器对行人目标分类。实验结果表明,采用改进过后的Stixel-world算法用于图像预处理极大地减少了计算时间。缩小了行人检测的候选区域,基于特征融合和hik-SVM的目标检测算法在保证检测准确度的前提下,具有较好的实时性和鲁棒性。
[Abstract]:In view of the problem that the pedestrian detection of monocular vision can not obtain depth information, which leads to more redundant information, the detection efficiency and accuracy are limited. In the stage of image preprocessing, an optimized analysis of parallax information generated by binocular stereo vision is proposed to simplify the expression of dynamic programming rod-like pixel scene in complex scene, and then, in the stage of pedestrian target detection, Based on the analysis of block scale in traditional HOG features and dimension reduction, the multi-scale hog multi-hog features suitable for road environment are selected by Fisher criterion, and the Multi-HOG features and LUV color channel features are fused. Finally, cross kernel support vector machine (SVM) classifier is used to classify pedestrian targets. The experimental results show that the improved Stixel-world algorithm can greatly reduce the computing time for image preprocessing. The candidate area of pedestrian detection is reduced and the target detection algorithm based on feature fusion and hik-SVM has better real-time and robustness on the premise of ensuring detection accuracy.
【作者单位】：湖南大学汽车车身先进设计制造国家重点实验室;
【基金】：国家自然科学基金(51475153) 深圳市科技计划(JCYJ20160530193357681)项目资助
【分类号】：TP391.41;U463.6

【相似文献】