الفهرس | Only 14 pages are availabe for public view |
Abstract 5.6 Summary and Conclusion For scene analysis given input RGB-D image, there are two processes running concurrently: Indoor scene recognition using deep learning to recognize the category of the RGB image, and RGB-D image segmentation. This image segmentation method includes five main steps: 1) the ‘edge detection’. It is applied on the depth image, 2) the ‘morphological operations’. Those operations are applied on the resultant image obtained from step 1, 3) the ‘connected component labeling’. It is applied on the negative of result of step 2, in order to assign label to the continuous regions, 4) is the extraction of the missing components and merging with result in step 3. In a similar way to the depth images, the previous four steps are applied on the RGB image, 5) combining the result of segmentation on the depth and the RGB images to form the final result. The segmentation accuracy obtained for this method is 73.41%. It takes an average time of 11sec and average storage of 83.8MB. The result represents separated objects in the given image. The objects are classified using multiple deep learning architecture (VGG-16 Net) for objects recognition based on the category of the scene of the given image. This process achieved a classification accuracy of 89.24%. Tracking recognized objects in image sequence in applied by extraction Shi Tomasi’s good features from each recognized object in the image. Then these features are tracked using pyramid Lucas and Kanade tracker. The proposed Tracking has improved the speed performance 10 times for a sequence of 12 image. The larger the image sequence the more improved speed performance. Finally, the semantic relations between objects are extracted using a predefined semantic rules. |