Abstract: Due to the influence of shape, illumination, visual collisions and visual blur, vehicle classification and counting are complex problems in video-based surveillance. Specifically, there is an indispensable part that should be handled carefully, i.e., the foreground extraction. In the initial foreground extraction, a model is established to determine whether there is a vehicle bonding. For the vehicles with visual collision, the model candetermine more accurate foregrounds through the gray space double threshold and the YCbCr image space processing. For such purpose, the paper defines the gap feature vector to recast the vehicle segmentation problem into the optimization problem of finding the dividing points, and gives an efficient vehicle segmentation algorithm to segment the vehicle. Final, a deep neural network is employed to classify the segmented vehicles. The experimental results show that our model is in a position to handle the surveillance videos in practice. Compared with manually calculating vehicle traffic or establishing a three-dimensional model to analyze vehicle classification and counting in the case of vehicle collisions, the proposed method takes the accuracy and timeliness into account. The efficiency is improved but the cost is reduced.

Key words: vehicle bonding, gray space, gap feature vector, segmentation, optimization, classification and counting

摘要： 由于受形态变化、光照变化、视觉碰撞和视觉模糊的影响,基于监控视频的车辆分类和计数一直都是待解决的复杂问题。为了更好地解决这个问题,本文提出新的模型来更好的提取前景。详细来讲,在初次前景提取中,建立模型判断是否存在车辆碰撞,对存在碰撞的车辆通过灰度空间双阀值和YCbCr图像空间处理后,对前景进行更准确的再提取。并在此基础上针对碰撞车辆,定义间隙特征向量将车辆分割问题转换为寻找分割点的优化问题,从而给出高效的车辆分割算法,对发生碰撞的车辆进行准确分割。之后利用神经网络对车辆分类,并设计一种基于已正确对碰撞车辆分割的算法对车辆计数。实验结果表明,本文提出的模型在视频车辆的分类和计数中取得优异的表现,并且数据处理速度能够满足及时性。比起人为计算车流量或建立三维模型等进行分析车辆碰撞情况下的车辆分类与计数,此方法兼顾了准确性与时效性,效率提高,成本减少。

关键词: 车辆碰撞, 灰度空间, 间隙特征向量, 分割, 优化, 分类和计数