【摘要】 监控视频受到现实场景、周边环境与拍摄设备的限制,所记录的信息会受到一定的噪声影响。论文以核相关滤波提取方法为基础,结合了现有的深度学习技术及跟踪方法,针对目标检测与跟踪算法中的特征表达和ID转换等难点问题进行了研究。结果显示该方法在针对目标漂移形变时具有良好的鲁棒性,且能很好地改进跟踪过程中的ID转换问题。更多还原

【Abstract】 The surveillance video is limited by the actual scene,the surrounding environment and the shooting equipment.The recorded information will be affected by certain noise. Based on the kernel correlation filtering extraction method,this paper combines the existing deep learning techniques and tracking methods to study the difficult problems such as feature representation and ID conversion in object detection and tracking algorithms. The results show that the proposed method has good robustness against object drift deformation and can improve the ID conversion problem in the tracking process.更多还原