融合CamShift和改进FairMOT下的Robei EDA安防视频树叶遮挡异常检测系统设计

Design of Robei EDA security video leaf occlusion anomaly detection system integrating CamShift and improved FairMOT

摘要:

安防视频场景多处于自然光不饱和条件下,光照强度、色温会随时间或天气产生动态波动,这种波动会直接干扰RGB色彩概率分布统计,导致亮度信息与色彩信息高度耦合,难以识别安防视频中的异常状况。为此,设计一种融合连续自适应均值漂移算法(Continuously Adaptive Mean-SHIFT,CamShift)色彩跟踪算法与改进多目标跟踪算法框架(Fair Multi-Object Tracking,FairMOT)的异常检测系统。基于Robei EDA平台构建硬件架构,采用FPGA、DSP与TVP5150编码器协同的三通道并行处理方案,FPGA负责视频采集与预处理,DSP执行目标检测与跟踪,TVP5150实现交互显示。以树叶遮挡区域的色彩信息为特征,将安防视频图像进行RGB色彩空间变换到HSV色彩空间,将树叶遮挡区域与背景分离,降低亮度信息与色彩信息的耦合性,利用CamShift对树叶遮挡区域进行自适应色彩跟踪。结合改进FairMOT的多特征融合相似度矩阵与重识别损失优化,采用重识别技术,通过重识别损失函数自适应识别叶子是否为同一叶子遮挡的新个体,判断检测目标是否为异常树叶遮挡目标,实现Robei EDA安防视频树叶遮挡异常检测。实验表明,该方法分区结果中共有5个非树叶网格,15个树叶网格,与实验指标一致。当FPR为10%时,TPR达94%,显著优于现有方法。并且可以满足如今安防监控系统中树叶遮挡检测问题的主要需求,具有可行性和应用价值。

Abstract:

Security video scenes are often under natural light saturation conditions, and the intensity and color temperature of the light will dynamically fluctuate with time or weather. This fluctuation will directly interfere with the probability distribution statistics of RGB colors, resulting in a high degree of coupling between brightness information and color information, making it difficult to identify abnormal situations in security videos. To this end, design an anomaly detection system that combines the Continuous Adaptive Mean Shift (CamShift) color tracking algorithm with an improved Multi Object Tracking (FairMOT) algorithm framework. Based on the Robei EDA platform, a hardware architecture is constructed using a three channel parallel processing scheme that combines FPGA, DSP, and TVP5150 encoder. FPGA is responsible for video capture and preprocessing, DSP performs object detection and tracking, and TVP5150 achieves interactive display. Using the color information of the leaf occlusion area as a feature, the security video image is transformed from the RGB color space to the HSV color space. The leaf occlusion area is separated from the background to reduce the coupling between brightness information and color information. CamShift is used for adaptive color tracking of the leaf occlusion area. Combining the improved FairMOT multi feature fusion similarity matrix with re identification loss optimization, re identification technology is adopted to adaptively identify whether a leaf is a new individual occluded by the same leaf through the re identification loss function, and determine whether the detection target is an abnormal leaf occluded target, achieving Robei EDA security video leaf occlusion anomaly detection. The experiment shows that there are a total of 5 non leaf grids and 15 leaf grids in the partitioning results of this method, which are consistent with the experimental indicators. When FPR is 10%, TPR reaches 94%, significantly better than existing methods. And it can meet the main needs of leaf occlusion detection in today"s security monitoring systems, with feasibility and application value.