Shadow detection in video surveillance by maximizing agreement between independent detectors

Juan Carlos San Miguel Avedillo and Jose M. Martinez
email: {JuanCarlos.SanMiguel, JoseM.Martinez} at uam.es

Video Processing and Understanding Lab (http://www-vpu.ii.uam.es/)
Escuela Politecnica Superior (http://www.ii.uam.es/)
Universidad Autonoma de Madrid (http://www.uam.es/)

Description

This work starts from the idea of automatically choosing the appropriate thresholds for the shadow detection. It is based on the maximization of the agreement between two independent shadow detectors without training data. Firstly, this shadow detection algorithm is described and then, it is adapted to analyze video surveillance sequences. Some modifications are introduced to increase its robustness in generic surveillance scenarios and to reduce its overall computational cost (critical in some video surveillance applications). Experimental results show that the proposed modifications increase the detection reliability as compared to some previous shadow detection algorithms and performs considerably well across a variety of multiple surveillance scenarios.

Video sequences

For each sequence, we provide the following results*:

• The original sequence
• The Foreground segmentation results using the described system in the paper.
• Shadow detection results on foreground segmentation (blue/red colour for foreground/shadow detection)
  • Standard HSV shadow detection algorithm [1]
  • Base shadow detection algorithm [2] (We want to thank the authors for provinding the source code)
  • Adaptation of [2] to process video surveillance sequences
  • Proposed algorithm with a fusion scheme based on the agreement of two (at least) detectors (a2)
  • Proposed algorithm with a fusion scheme based on the agreement of the three detectors used (a3)
  • Comparative between the following algorithms:
    • Base algorithm [2], named "DCU" in the comparative video sequence (upper left)
    • Adaptation of the Base algorithm, named "DCU Adapted" in the comparative video sequence (upper right)
    • Proposed algorithm with a fusion scheme based on the agreement of two detectors, named "VPU2" in the comparative video sequence (lower left)
    • Proposed algorithm with a fusion scheme based on the agreement of three detectors, named "VPU3" in the comparative video sequence (lower right)

                     

Dataset	Sequence	Foreground Segmentation	Shadow detection results
			Standard HSV	Base alg.	Base alg. adapted	Proposed alg.  Agr. 2	Proposed alg.  Agr. 3	Comparative
PETS 2006	S1_T1_C3	XVID	XVID	XVID	XVID	XVID	XVID	XVID
	S4_T5_A3	XVID	XVID	XVID	XVID	XVID	XVID	XVID
AVSS 2007	AB_Easy	XVID	XVID	XVID	XVID	XVID	XVID	XVID
	PV_Easy	XVID	XVID	XVID	XVID	XVID	XVID	XVID
ATON	Campus	XVID	XVID	XVID	XVID	XVID	XVID	XVID
	Int. Room	XVID	XVID	XVID	XVID	XVID	XVID	XVID

 *The xvid codec is needed to watch the video files (download it here)    

References

[1] Cucchiara, R.; Grana,C.; Piccardi,M.; Prati,A. “Detecting moving objects, ghosts and shadows in video streams”, IEEE Trans. on Pattern Analysis and Machine Intelligence, 25(10):1337-1342, 2003.

[2] Conaire, C.O.; O'Connor, N.E.; Smeaton, A.F., "Detector adaptation by maximising agreement between independent data sources," Proc of IEEE Conf. on Computer Vision and Pattern Recognition, pp.1-6, 2007