基于有限Radon特征和判别稀疏字典学习的行人检测算法研究

This project focuses on computer vision based solutions for the pedestrian detection problem which has significant research importance and wide application potential. The Research work is conducted in both aspects of feature extraction and classifier design to address the critical issues in both the theory and application of robust pedestrian detection with high localization accuracy under complex environments suffering from complex illumination conditions and frequent occlusion. This project aims to establish a pedestrian detection system which is more coincidence with human cognitive character, and improve the robustness of pedestrian detection and the accuracy of the positioning. A new global feature which can provide effective joint contour and texture description is proposed, which computes the statistics of local finite Radon transform coefficients within overlapped blocks in candidate sliding windows. It is expected to have good noise and occlusion insensitivity. Considering the sparsity of natural image encoded by the primary visual cortex of the visual system, discriminant sparse dictionary classifier is adopted as the classifier. The decision-making output modules are improved based on Bayesian inference theory and thus it is expected to be effective to reduce the interference of the partial occlusion and improve the detection performance. The fusion mechanism of multi-resolution and multi-modal information is also researched under the proposed unified output probability space. The confidence of each candidate sliding window containing a pedestrian target is given which is expected to solve the robust pedestrian detection problem in complex illumination conditions and frequent occlusion in the real application.

针对具有重要理论意义和广泛应用价值的行人检测问题，本课题基于计算机视觉方法，从特征提取及分类器设计两个方面探讨鲁棒检测理论及实际应用中复杂光照、频繁遮挡等棘手问题的解决方案，建立更加符合人类认知特性的行人检测系统，提高行人检测鲁棒性和定位准确性。特征提取方面，拟采用有限Radon域局部特征，计算其在候选滑动窗口中相互重叠块内局部特征统计信息以构建适用于静止图像的行人检测的整体特征，预期可明显提高对轮廓及纹理信息的联合表示能力及对噪声、光照变化等因素的鲁棒性；分类器设计方面，基于视觉系统初级视皮层对自然图像的编码稀疏性，拟采用判别稀疏字典分类器，并基于贝叶斯理论对其决策输出模块进行改进，预期可有效降低部分遮挡的干扰。在上述改进算法给出的统一输出概率空间下，研究多分辨率、多模态信息的融合机制，给出候选滑动窗口内包含行人目标的的置信度，解决光照复杂、频繁遮挡等实际应用环境下的鲁棒行人检测问题。

基于计算机视觉方法的行人检测问题具有重要理论意义和广泛应用价值，经过三年的研究，本课题在行人检测过程中的特征提取及分类器设计两个方面取得了一定的研究成果，建立了更加符合人类认知特性的行人检测系统，提高了行人检测鲁棒性和定位准确性。特征提取方面，提出了基于有限Radon域的行人局部特征表示方法，计算其在候选滑动窗口中相互重叠块内局部特征统计信息以构建适用于静止图像的行人检测的整体特征，明显提高对轮廓及纹理信息的联合表示能力及对噪声、光照变化等因素的鲁棒性；分类器设计方面，提出了采用判别稀疏字典的行人检测分类器，并基于贝叶斯理论对其决策输出模块进行改进，有效降低了存在部分遮挡时的干扰。在上述改进算法给出的统一输出概率空间下，研究了多分辨率、多模态信息的融合机制，给出候选滑动窗口内包含行人目标的的置信度，提升了光照复杂、频繁遮挡等实际应用环境下的鲁棒行人检测问题的准确性和可靠性。