-
为了验证方法的有效性和降噪指标的合理性,分别用仿真数据和真实场景数据进行对比实验分析。此外,为对比降噪效果,将Inivation公司开发的DV软件中开发的Dvsnoisefilter模块和文中算法对比,该滤波器主要参数如表1所示。
Parameter Value Background activity support min 1 Background activity support max 8 Background activity time/μs 2 000 Refractory period time/μs 100 Table 1. Main parameters of Dvsnoisefilter
-
事件相机仿真器[13]是模拟事件相机生成数据的软件,使用该仿真器将一段高帧视频转化为模拟事件流数据。高帧视频是在相机固定情况下拍摄单个行人获得的,如图3(a)所示。由于相机处于固定状态,场景运动目标单一且简单,仿真得到的事件流可看作无噪声事件流数据,图3(b)为仿真事件3D可视化效果,图3(c)为2D可视化效果。
Khodamoradi团队[14]证实了事件相机的噪声事件,也称背景活动(Background Activity,BA)[15],可用泊松分布来描述:
式中:t为时间间隔;n为该时间段内的BA事件数;λ为每个像素点产生BA的平均速率。
为了验证算法的效果,首先在仿真事件流数据中分别加入一定比例(10%、20%、50%和100%)、符合泊松分布的噪声事件,图4(a)为加入噪声后的可视化效果图。其次,分别用滤波法和文中算法对上述加入噪声后的数据集进行去噪,如图4(b)和图4(c)所示。最后,通过对比去噪前后二维可视化效果图定性分析去噪效果。此外,为了对去噪算法定量分析,统计并对比了滤波法和文中算法真实事件率和事件信噪比。实验结果如图3~图5和表2~表3所示。
Method Real events rate, Ptrue 10% 20% 50% 100% Initial state 90.90% 83.33% 66.67% 50.00% Filtering 98.03% 97.31% 96.54% 89.01% Proposed 98.66% 98.21% 97.59% 93.89% Table 2. Statistics table of true event rate under different noise
结果表明,随着噪声比率的增加,滤波法降噪后目标周围出现越来越多的噪声事件未被过滤掉,出现这种情况和阈值大小设置有关系,而文中算法只是人内部和周围极少噪声事件被当作有效。另外随着噪声比率增加,真实事件率和事件信噪比都有所下降,但是文中算法下降较小,尤其是在噪声较大的情况下,文中算法相对于滤波法的优势更明显,从而也证明了文中算法的有效性。
-
在该实验中,真实场景数据来源事件相机DAVIS346,其空间分辨率为346×260,动态范围120 dB,延迟20 μs,芯片功率10~170 mW。
为了验证算法的有效性,使用DAVIS346捕获的室内的场景事件流数据作为验证数据集。该场景背景复杂,包含桌子、饮水机、柜子、台灯等物体,帧图像如图6(a)所示。真实数据3D、2D可视化图像如图6(b)和图6(c)所示,从可视化效果图中可以看出相机拍摄数据中存在大量的噪声。
使用滤波法和文中算法对采集的数据进行降噪,降噪后3D和2D可视化效果如图7所示,其中图7(a)和图7(b)为滤波法降噪效果,图7(c)和图7(d)为文中算法降噪效果。从可视化效果上分析,文中算法更好地去除了目标边缘处的噪声,使边缘轮廓更加清晰,明显更优于滤波方法。此外,滤波法降噪后剩余35524个事件,文中算法降噪后剩余32982个事件,降噪精度高于滤波法,见表4。文中算法直接对事件流数据降噪,利用相机高时间分辨率的特点设计的去噪算法可有效去除噪声,去噪后数据可直接应用于后续的事件流处理算法中。
Method Loss rate of real events, Ploss 10% 20% 50% 100% Filtering 1.87% 2.69% 3.84% 4.25% Proposed 1.75% 2.51% 3.44% 3.80% Table 3. Statistics table of true event loss rate under different noise
Number Pprecision Method 50000 - Filtering 35524 28.95% Proposed 32982 34.04% Table 4. Statistics of noise reduction accuracy of real events by different algorithms
Denoising algorithm based on improved Markov random field for event camera
doi: 10.3788/IRLA20210294
- Received Date: 2021-05-05
- Rev Recd Date: 2021-06-21
- Available Online: 2021-10-20
- Publish Date: 2021-10-20
-
Key words:
- event camera /
- denoising algorithm /
- probability undirected graph model /
- conditional iteration mode
Abstract: To solve the problem of the large amount of noise in the event stream output by the event camera, an event stream denoising algorithm based on the probability undirected graph model was introduced. Due to the imaging principle of the camera, the change of the target had certain regularity and correlation in time and space. By mapping the event to the polar coordinate space-time neighborhood, the local correlation of the event was established to build a complete probability graph model. In addition, the improved conditional iterative mode algorithm was used to optimize the iterative solution of model. The experimental results of simulated data generated by the event camera simulator and the real data recorded by DAVIS346 show that the proposed algorithm can effectively remove noise events. Finally, the comparison with the filtering algorithm proves that the algorithm is superior to the filtering algorithm.