【摘要】 通过Opencv调用设备摄像头将数据读入,视频流切分成图片流后进入在Tensorflow框架下训练好的FasterR-CNN模型进行目标识别,将识别出的目标进行标记并输出。在教室环境下由于相互遮挡和密集度大的问题严重,原始的训练模型不能很好地识别困难目标,通过收集大量教室监控截图并进行Data Augmentation数据增强来进一步地扩大样本,提高识别准确度;使用LabelImg对数据进行标注生成训练和测试数据集,将VOC2007及新标注样本进行合并得到NEW_VOC2007,在此新数据集上进行训练及测试生成新模型。实验证明,新模型能够更好地适应教室环境的识别,准确度更高。更多还原

【Abstract】 In the school environment,pedestrian-intensive,sheltered,multi-scale problems are common.In this paper,Faster R-CNN is used for pedestrian detectionin complex environment of school buildings.When training samples,A training strategy for Difficult Sample Mining is introduced,which adjusts the weight of difficult samples while picking out difficult samples,so as to make training more focused.LabelImg is used to lable the data sampled in complex environment.Then VOC2007 and the annotated samples are merged to obtain extended VOC2007 data set.Training and testing are carried out on this basis to establish the model with good performance.The experimental results show that compared with the four-step training method commonly used in Faster R-CNN,the generalization performance is improved by using this training method.更多还原