为提升辅助驾驶系统的可靠性及安全系数, 实现更高精度的行人检测, 基于人体树图模型提出了一种改进的离线训练、在线检测的行人检测方法。首先, 定义人体部件间的共生关系, 得到对应父子部件对, 结合K-means算法对其位置关系进行聚类获得部件类型。为兼顾类内紧密性与类间分离性, 采用MSE和DBI构建具有两阶段适应度函数的混合粒子群聚类算法, 在有效估计各部件最优聚类中心数量的同时, 消除随机初始化对聚类准确率造成的影响。其次, 将优化聚类得到的部件类型作为隐藏变量, 通过求解隐结构SVM获取改进后的人体检测模型。最后, 通过动态规划算法求解状态转移方程, 在多个尺度上有效估计人体部件位置及检测包围盒, 并结合非极大值抑制思想得到最终的行人检测结果。实验结果表明, 所提方法在检测性能上明显优于5种行人检测方法, 并且相比于原始Pose-original方法, 在INRIA和ETH数据集上的丢失率分别下降了8.14%和5.05%。实验证明该方法检测性能良好且具有较高的准确性和鲁棒性。

In order to improve the reliability and safety factor of driver assistance systems, and achieve pedestrian detection with a higher precision, an improved pedestrian detection method based on a tree-structured graphical model of the human body is proposed, and it consists of an offline training part and an online detection part. First, the corresponding parent-child parts are obtained by defining the symbiotic relationship between human parts, and then the K-means algorithm is applied to the location relationship between part pairs to acquire part types via clustering. For the purpose of taking both intra-class tightness and inter-class differences into account, a hybrid particle swarm optimization algorithm is built with a two-phase fitness function via introducing MSE and DBI. It is not only effective in estimating the number of optimal cluster centers, but also in eliminating the effect of random initialization on the clustering accuracy. Then, the part type obtained using the optimized clustering method is considered as the latent variable. The pedestrian detection model is obtained through solving the latent structural SVM problem. Finally, we estimate the position of human parts and the detection bounding box on multiple scales based on solving the state equation via a dynamic programming algorithm, and obtain the final pedestrian detection result through incorporating the idea of non-maximum suppression. Experimental results indicate that the performance of the proposed algorithm is superior to those of five other pedestrian detection algorithms. In particular, on the INRIA and ETH databases, the loss rate of the proposed algorithm decreased by 8.14% and 5.05%, respectively, compared with that of the pose-original method. Experimental results show that the proposed algorithm has good performance and high accuracy and robustness.