Comparison of Two Still Image Wearable Camera Annotation Schemas for Posture Classification
Mentor 1
Scott Strath
Start Date
28-4-2023 12:00 AM
Description
Wearable cameras (WC) are tools that provide rich contextual information on physical activity. WC still image annotation coding schemas are typically developed to answer study-specific research hypotheses, which limits standardization and comparability of results. To compare two WC image annotation schemas to identify posture behavior. 25 healthy adults were recruited to wear an Autographer (OMG Life) WC for a full day (>600 mins). Still images attained from the full day were coded for posture behavior on two different annotation schemas, from the University of Wisconsin-Milwaukee (UWM) and from Oxford University (OXF). All coders were extensively trained on both schemas and achieved inter-rater reliability scores of >0.8 on postural classifications in training data sets. The UWM schema has a separate set of codes for posture classification. The OXF schema is based off the Compendium of Physical Activities where codes are posture dependent. For analysis, sitting, standing, and movement classifications were directly compared between the two schemas. Second by second percent agreement of schemas was utilized to determine how well the applied codes matched between schemas. A confusion matrix was created to determine how well the schemas agreed/disagreed on the posture codes applied. 10583.1 minutes (176 hours) were analyzed. Overall percent agreement between the UWM and OXF schemas was 93.7%. By each posture classification, the schemas had a 97.81% agreement for sitting, 63.48% for standing, and 83.96% for movement. WC still image schemas can provide comparable results for time spent in different postural behaviors. Future research on this topic should examine agreement on other key features of physical behavior type.
Comparison of Two Still Image Wearable Camera Annotation Schemas for Posture Classification
Wearable cameras (WC) are tools that provide rich contextual information on physical activity. WC still image annotation coding schemas are typically developed to answer study-specific research hypotheses, which limits standardization and comparability of results. To compare two WC image annotation schemas to identify posture behavior. 25 healthy adults were recruited to wear an Autographer (OMG Life) WC for a full day (>600 mins). Still images attained from the full day were coded for posture behavior on two different annotation schemas, from the University of Wisconsin-Milwaukee (UWM) and from Oxford University (OXF). All coders were extensively trained on both schemas and achieved inter-rater reliability scores of >0.8 on postural classifications in training data sets. The UWM schema has a separate set of codes for posture classification. The OXF schema is based off the Compendium of Physical Activities where codes are posture dependent. For analysis, sitting, standing, and movement classifications were directly compared between the two schemas. Second by second percent agreement of schemas was utilized to determine how well the applied codes matched between schemas. A confusion matrix was created to determine how well the schemas agreed/disagreed on the posture codes applied. 10583.1 minutes (176 hours) were analyzed. Overall percent agreement between the UWM and OXF schemas was 93.7%. By each posture classification, the schemas had a 97.81% agreement for sitting, 63.48% for standing, and 83.96% for movement. WC still image schemas can provide comparable results for time spent in different postural behaviors. Future research on this topic should examine agreement on other key features of physical behavior type.
