基于多目标聚类的车辆检测方法（英文）

Multi-objective Clustering Based Approach for Vehicle Detection

为提升车辆检测算法中字典学习的有效性,提出一种新的基于多目标聚类的车辆检测方法。同时考虑聚类检测中的全局偏差和连接性2个重要的指标,并引入提出的新多目标优化方法,期望获得1组同时符合全局偏差和连接性的平衡解。针对字典学习和多目标聚类解的特性,设计了一种新的模型选取算法,用于选出有利于提高检测性能的最终聚类解。对5个车辆部件的聚类任务进行聚类,以验证所提算法的有效性与卓越性。将该方法与5种其他方法(不仅包含2类常用单目标聚类方法,也覆盖了其他多目标聚类方法)对2类车辆检测问题进行检测,以全面检验该方法的性能。结果表明:双目标聚类在车辆检测应用中,对聚类效果有较好促进作用,且该方法的整体性能相比于其他聚类算法,颇具竞争性。

To promote the effectiveness of codebook learning in vehicle detection algorithm, a new multi-objective clustering based approach for vehicle detection was proposed. Two key indicators in clustering detection, overall deviation and connectedness, were considered simultaneously, and a novel multi-objective optimization algorithm was introduced, which expected to obtain a set of balanced solutions between overall deviation and connectedness. In order to get a final clustering solution, which is beneficial for improvement of the detection performance, a new model selection algorithm was designed according to the characteristics of codebook learning and multi-objective clustering solution. Clustering was carried out on 5 clustering tasks for vehicle parts to verify its effectiveness and superiority. Besides, two kinds of vehicle detection problems were detected by the proposed method and five other methods which involved single objective as well as other multi-objective clustering methods to comprehensively examine the performance of the method. The results show that in the application of multi-objective clustering to vehicle detection, the clustering effect is greatly promoted, and the overall performance of the proposed method is very competitive, comparing to other clustering algorithms.