Academia.eduAcademia.edu

Figure 2 – uploaded by Ilkay Ulusoy

See full PDFdownloadDownload figure
Table 4 Prevision and recall rates of the proposed method for the dataset sequences

Table 4 Prevision and recall rates of the proposed method for the dataset sequences

Related Figures (12)

Table 1 Precision and recall rates of the proposed algorithm for the OSU dataset when different background models are applied
Table 1 Precision and recall rates of the proposed algorithm for the OSU dataset when different background models are applied
Fig. 2 Flowchart of the algorithm
Fig. 2 Flowchart of the algorithm
Finally, the shadow image is masked with the foreground image and the actual shadow regions are extracted (Fig. 4e). The presence of shadows is one of the main problems for the visible domain. Before the fusion of the foreground regions of the visible domain into the infrared domain, shadows are removed from the colour foreground image. Three methods of shadow detection are implemented and compared. The first method assumes that the difference between the ratios of the normalised components of each channel of the current pixel and the background model should be lower
Finally, the shadow image is masked with the foreground image and the actual shadow regions are extracted (Fig. 4e). The presence of shadows is one of the main problems for the visible domain. Before the fusion of the foreground regions of the visible domain into the infrared domain, shadows are removed from the colour foreground image. Three methods of shadow detection are implemented and compared. The first method assumes that the difference between the ratios of the normalised components of each channel of the current pixel and the background model should be lower
moved towards its interior normal and stopped on the boundary of the object. Equation (8) explains the main idea These shadow regions are removed from the colour image before fusing the two domains.
moved towards its interior normal and stopped on the boundary of the object. Equation (8) explains the main idea These shadow regions are removed from the colour image before fusing the two domains.
Fig.5 Infrared (a), intensity (b) and colour (c) images are given in the top row. The foreground detection results are given for infrared only (d, g, j), colour only (e, h, k) and fused (f, i, 1) cases in the first, second and third columns, respectively. The SG (d, e, f), the non-parametric (g, h, i) and the MOG (j, k, 1) results are given in the second, third and fourth rows, respectively
Fig.5 Infrared (a), intensity (b) and colour (c) images are given in the top row. The foreground detection results are given for infrared only (d, g, j), colour only (e, h, k) and fused (f, i, 1) cases in the first, second and third columns, respectively. The SG (d, e, f), the non-parametric (g, h, i) and the MOG (j, k, 1) results are given in the second, third and fourth rows, respectively
Table 3 Overall precision and recall rates of the proposed method and five other methods for the OSU dataset
Table 3 Overall precision and recall rates of the proposed method and five other methods for the OSU dataset
Table 2 Precision and recall rates of the proposed method and [1] for the OSU dataset sequences
Table 2 Precision and recall rates of the proposed method and [1] for the OSU dataset sequences
Fig. 7 Results of frames from the dataset when different background modelling algorithms are applied
Fig. 7 Results of frames from the dataset when different background modelling algorithms are applied
post-processing such as clearing wrong contours or completing missing contour sections is performed. Instead, the contour information is obtained immediately after the snake algorithm. unfair to compare our method with our implementation of their method. Table 2 gives the precision and recall rates for the proposed method, for [1] and then for [1] when only the infrared domain was used. In all OSU sequences, the recall values of the proposed method are higher but the precision values are lower. Higher recall implies that more of the detected pixels are actual object pixels. Since, the precision of the fused method is always less than when only infrared data are used, the expected outcome of fusion is an increase in recall and this is achieved by the proposed algorithm.
post-processing such as clearing wrong contours or completing missing contour sections is performed. Instead, the contour information is obtained immediately after the snake algorithm. unfair to compare our method with our implementation of their method. Table 2 gives the precision and recall rates for the proposed method, for [1] and then for [1] when only the infrared domain was used. In all OSU sequences, the recall values of the proposed method are higher but the precision values are lower. Higher recall implies that more of the detected pixels are actual object pixels. Since, the precision of the fused method is always less than when only infrared data are used, the expected outcome of fusion is an increase in recall and this is achieved by the proposed algorithm.
Fig.8 Fused foreground mask and detected foreground objects for frames of the dataset
Fig.8 Fused foreground mask and detected foreground objects for frames of the dataset
In future, the method will be applied to the detection and identification of other objects. By using some basic parameters such as the area of the blob and height to width ratio, object classification will be performed. Moreover, further tasks such as activity recognition from human silhouettes can be undertaken based on the method outlined in this paper. arrect the background model in infrared domain. Ihe temperature of the background is more stable over time and operation over 24h is valid in the infrared domain. These are the main reasons that the initial regions of interest are first found in the infrared domain, and these regions are used as a mask for the visible domain. However, some object regions may not be detected from the infrared image because of the temperature of the foreground objects being very close to the background model. Therefore after finding the corresponding foreground regions in the visible domain, the visible cues (if available) are added to the foreground regions of the infrared domain. As a result, the missing regions on the thermal foreground mask are completed using the corresponding visual information. Finally, application of the snake algorithm to the current infrared image, which is masked with the final foreground regions, gives very successful results for people detection. The results are more robust when the halo effects surrounding the people are more evident.
In future, the method will be applied to the detection and identification of other objects. By using some basic parameters such as the area of the blob and height to width ratio, object classification will be performed. Moreover, further tasks such as activity recognition from human silhouettes can be undertaken based on the method outlined in this paper. arrect the background model in infrared domain. Ihe temperature of the background is more stable over time and operation over 24h is valid in the infrared domain. These are the main reasons that the initial regions of interest are first found in the infrared domain, and these regions are used as a mask for the visible domain. However, some object regions may not be detected from the infrared image because of the temperature of the foreground objects being very close to the background model. Therefore after finding the corresponding foreground regions in the visible domain, the visible cues (if available) are added to the foreground regions of the infrared domain. As a result, the missing regions on the thermal foreground mask are completed using the corresponding visual information. Finally, application of the snake algorithm to the current infrared image, which is masked with the final foreground regions, gives very successful results for people detection. The results are more robust when the halo effects surrounding the people are more evident.
Related topics:
Computer ScienceArtificial IntelligenceComputer VisionElectrical and Electronic Engineeringrecall rate
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved