《Real-time vehicle detection and counting in complex traffic scenes using background subtraction model with low-rank decomposition》要點總結
1、研究目的
1)To efficiently collect the real-time traffic information.
2)To efficiently overcome the main weaknesses of BS.One major drawback of BS is that modelling the background in real scenarios is not easy.
2、創新點
1)We online detect moving objects by the ‘low-rank + sparse’ decomposition of an input matrix Z = L + E, where sparse outlier E consists of two parts, moving objects S and noise G, E = S + G.
A typical noise G may be arisen from various illumination conditions or waving trees.
2)Group sparsity constraint is used to further improve the accuracy of foreground detections in complex dynamic scene. The group sparsity constraint, taking the spatial-temporal structure of neighbourhood pixels into consideration, makes the algorithm more robust to random noise.
3)Traffic objects detection is implemented in an online manner without requiring the entire image sequences. So our method can online update the background model and is robust to dynamic sceneries.
3、研究方法
Methods under this framework follow the idea that background sequence is modelled by a low-rank matrix and the moving objects correspond to the sparse outliers.
In this paper, the proposed BS method is based on two constraints. The first one is the background areas in video sequence are linearly correlated with each other, i.e. areas will arise a low-rank matrix. The second one is that we have the prior knowledge that, one moving object is observed by a sparse and contiguous piece in consecutive frames, and usually occupies a relatively small
portion of an image. In addition, locations of a moving object in successive frames are likely to group together, instead of scattered randomly. This indicates that the moving objects satisfy the group sparsity constraint .
1)Online background modelling
2)Online foreground detection
3)Real-time vehicle counting
4、實驗
Experiments are done on several traffic video sequences.
The video sequences are available in’ChangeDtetection.net’(CDnet2014) and Youtube.
The proposed method is tested on these traffic video datasets and compared with the state-of-the-art algorithms, including MOG, GoDec, Decolor, Corola and incPCP.
For quantitative evaluation, the metrics of precision, recall and the F-measure are used to show the overall accuracy.
5、進一步研究
6、相關知識
