融合聚类法的改进三帧差分车辆检测算法

Improved Three-Frame Differential Vehicle Detection Algorithm Incorporating Clustering Algorithm

针对三帧差分法在车辆检测任务中出现的前景点误检和漏检问题,提出了一种融合K-means聚类的改进三帧差分车辆检测算法。首先,综合当前图像分别与改进算法所选两帧的差分结果,初步判定像素点类别并定义待分类点;其次,结合待分类点在三帧内的灰度特征对其进行K-means聚类,并依据点的坐标信息修正聚类结果,得到待分类点类别;最后,设计车辆形状修正方法,填补空洞并修正目标边界,完成检测。实验结果显示,改进算法在2种不同场景视频上的检测效果达到了81.72%的平均精确率、93.85%的平均召回率以及87.34的平均F1值,各指标值相比于原三帧差分法平均有11.86%提升,较好解决了检测中前景点误检和漏检的问题。

Aiming at the problem of false detection and missed detection of foreground points in vehicle detection tasks by the three-frame differential method, an improved three-frame differential vehicle detection algorithm incorporating K-means clustering is proposed. Firstly, the difference results between the current image and the two frames selected by improved algorithm are combined to initially determine the category of pixel points and define the points to be classified. Secondly, the K-means clustering of the points to be classified by combining their grayscale features within three frames and correct the clustering results based on the coordinate information of the points, and the category of the points to be classified is obtained. Finally, the vehicle shape correction method is designed to fill the holes and correct the target boundary, and the detection is completed. The experimental results show that the detection effect of improved algorithm on video of two different scenarios reaches an average precision of 81.72%, an average recall rate of 93.85%, and an average F1value of 87.34%. Compared with the three-frame differential algorithm, the improved algorithm shows an average improvement of 11.86% in each metric. This proves that the improved algorithm well solves the problem of false and missed detection of foreground points in the detection.