基于图像序列的运动目标检测识别的关键技术研究

为提高运动目标检测的识别效果,通过分析、综合比较各种运动目标检测算法的优劣性,提出了基于全局自适应帧差法和基于码本模型的背景减除法对同一运动目标进行检测。通过对运动目标检测提取运动目标的掩膜,对掩膜进行外接矩形分析,从而得到包围运动目标的矩形框;将矩形框内的图片截取出来,调整该矩形并提取图片的HOG特征,最后通过训练好的SVM进行分类。在训练过程中,针对难易情况应用自举法对训练器进行优化。实验表明,与传统HOG+SVM多尺度检测算法相比,该方法在速度和准确性上可提升20%左右,可作为运动目标检测与识别的参考方法。

Research on Key Techniques of Moving Target Detection and Recognition Based on Image Sequence

To improve the recognition of moving target detection, the pros and cons of various moving target detection algorithms were analyzed, the global adaptive frame difference method and the codebook model based background subtraction method was proposed to detect moving targets simultaneously for the same target.The method used the masks of moving target and then analyzed them by rectangles, so that the pictures in the rectangle can be extracted out.Next, adjust the rectangle and extract the HOG features of the image, classify through trained SVM.In the training process, the bootstrap method is applied to optimize the trainer for difficult situations.Experiments show that compared with the traditional HOG+SVM multiscale detection algorithm, this method can increase the speed and accuracy by about 20%, and it can be used as a reference method for moving target detection and recognition.