|dc.description.abstract||Popular video sharing websites contain a large collection of videos in various sign languages. These websites have the potential of being a significant source of knowledge sharing and communication for the members of the deaf and hard-of-hearing community. However, prior studies have shown that traditional keyword-based search does not do a good job of discovering these videos.
Dr. Frank Shipman and others have been working towards building a distributed digital library by indexing the sign language videos available online. This system employs an automatic detector, based on visual features extracted from the video, for filtering out non-sign language content. Features such as the amount and location of hand movements, symmetry of motion etc. have been experimented with for this purpose. Caio Monteiro and his team designed a classifier which uses face detection to identify the region-of-interest (ROI) in a frame, and foreground segmentation to estimate amount of hand motion within the region. It was later improved upon by Karappa et al. by dividing the ROI using polar coordinates and estimating motion in each division to form a composite feature set.
This thesis work examines another visual feature associated with the signing activity i.e. speed of hand movements. Speed based features performed better compared to the foreground-based features for a complex dataset of SL and non-SL videos. The F1 score showed a jump from 0.73 to 0.78. However, for a second dataset consisting of videos with single signers and static backgrounds, the classification scores dipped. More consistent performance improvements were observed when features from the two feature sets were used in conjunction. F1 score of 0.76 was observed for the complex dataset. For the second dataset, the F1 score changed from 0.85 to 0.86.
Another associated problem is identifying the sign language in a video. The impact of speed of motion on the problem of classifying American Sign Language versus British Sign Language was found to be minimal. We concluded that it is the location of motion which influences this problem more than either the speed or the amount of motion.
Non-speed related analyses of sign language detection were also explored. Since the American Sign Language alphabet is one-handed, it was expected that videos with left-handed signing might be falsely identified as British Sign Language, which has a two-handed alphabet. We briefly studied this issue with respect to our corpus of ASL and BSL videos and discovered that our classifier design does not suffer from this issue. Apart from this, we explored speeding up the classification process by computing symmetry of motion in the ROI on selected keyframes as a single feature for classification. The resulting feature extraction was significantly faster but the precision and recall values depreciated to 59% and 62% respectively for a F1 score of .61.||en