改进星型级联可形变部件模型的行人检测

发布时间：2018-06-20 23:27

本文选题：星型级联检测可形变部件模型 + 行人检测　；参考：《中国图象图形学报》2017年02期

【摘要】：目的行人检测是计算机视觉和模式识别领域的研究热点与难点,由于经典的可形变部件模型(DPM)检测速度太慢,引入PCA降维的星型级联检测可形变部件模型(casDPM)相比较于DPM模型检测速度虽然有了很大提升,但在应用于行人检测时,出现检测精度较低、平均对数漏检率较高的情况,为了更加准确地对行人进行检测,提出了一种改进casDPM模型的行人检测方法。方法首先利用对象度量方法获取目标候选区域,结合目标得分信息得到casDPM模型低分检测区域的置信度,在设定的阈值上保留检测窗口;然后针对casDPM模型原有非极大值抑制(Nms)算法只利用单一的面积信息,造成误检数较高的情况,提出了利用检测窗口的得分信息进行改进;最后将两种方法结合起来,提出了融合的cas-WNms-BING模型。结果采用本文方法在INRIA数据集上进行检测,实验结果表明该方法对于行人形变、背景特征复杂及遮挡现象具有较强的鲁棒性,相比casDPM模型,本文提出的方法平均精度(AP)可以提高1.74%,平均对数漏检率可以降低4.45%。结论提出一种改进星型级联可形变部件模型,取得一定的研究成果,在复杂的背景下,能够有效地进行行人检测,主观视觉感受和客观实验评价指标都表明该方法可以有效提升模型行人检测效果。但是,星型级联可形变部件模型训练及检测效率仍有待提高,需进一步对模型存在的一些局限性进行深入研究。
[Abstract]:Objective pedestrian detection is a hot and difficult point in the field of computer vision and pattern recognition, because the detection speed of the classical deformable component model (DPM) is too slow. Compared with DPM model, PCA reduced dimension star cascade detection model cascase DPMhas a higher detection speed than DPM model, but when it is applied to pedestrian detection, the detection accuracy is lower and the average logarithmic leakage rate is higher. In order to detect pedestrians more accurately, an improved casDPM model is proposed for pedestrian detection. Methods firstly, the target candidate region was obtained by using object metric method, and the confidence of low score detection area of casDPM model was obtained by combining target score information, and the detection window was retained on the set threshold. Then aiming at the case where the original non-maximum suppression (Nms) algorithm of casDPM model only uses a single area information, which results in a higher number of false detections, the paper proposes to improve the algorithm by using the score information of the detection window. Finally, the two methods are combined together. A fusion cas-WNms-BING model is proposed. Results the method is used to detect the data set of INRIA. The experimental results show that the method is robust to pedestrian deformation, complex background features and occlusion, compared with casDPM model. The average precision of the method presented in this paper can be increased by 1.74 and the average logarithmic leakage rate can be reduced by 4.45%. Conclusion an improved star cascade deformable component model is proposed, and some research results are obtained, which can effectively detect pedestrian in complex background. Both subjective visual perception and objective experimental evaluation index show that this method can effectively improve the pedestrian detection effect of the model. However, the training and detection efficiency of star cascade deformable components model needs to be improved, and some limitations of the model need to be further studied.
【作者单位】：南京理工大学计算机科学与工程学院;苏州吴江区公安局;公安部第三研究所;
【基金】：国家自然科学基金项目(61371168) 江苏省科技支撑项目(BE2014646) 苏州市科技支撑项目(SS201413)~~
【分类号】：TP391.41

【参考文献】