Here is a brief overview of three approaches to analyzing mobile eye tracking output.

1. Watch the video output and manually tally the ‘hits’. Slowly playback the video and count how often the point of gaze lands on the area of interest. This gives you a general sense of what is looked at and what is ignored, but only limited conclusions can be drawn with respect to the amount of time spent viewing each AOI. One could slow the video down further (frame by frame) and take note of the time stamp for each time the participant starts looking at one particular region. This is enormously tedious, especially for video segments longer than 2 minutes and with more than just a handful of AOI. It is also imprecise; the time stamps are obtained at 30 fps, with up to 5 or more data points lost with each frame, before and after the time stamp. There is also substantial opportunity for human error in tallying hits or recording time (I speak from experience).

2.Identify fixations, then tag AOI based on fixation. In this approach the analysis software identifies fixations following a particular algorithm developed by the manufacturer. Then, thumbnails are generated for the video segment during which the fixation occurred. The research analyst tags each thumbnail with an identification of which AOI is being looked at during that fixation. The output sums the tags and can compute dwell time based on the length of the fixation. This is all very promising and most certainly not as time consuming as option 1. Nonetheless, each data file needs to be addressed individually and to a certain degree, manually. A greater concern, in my opinion, is the reliance on the identification of fixations to then in turn make conclusions about viewing attention and dwell time. Fixations, by definition, assume a moment of movement cessation of the pupil as the eye fixates on an object. What if the object is moving and the pupil is following in smooth pursuit? How is the fixation captured? Experience with this approach has left me wondering why the dwell times on AOIs total up to only a fraction of the total testing session. What did the participant look at the rest of the time? Did he keep moving his eyes so quickly that he never truly looked at anything? Hard to imagine.

3. Draw regions ‘on’ the output videos and process the data against the defined AOI. This is essentially the same idea as most analysis approaches with desktop systems. Identify the AOI in the scene and the software will tally when the x,y coordinates of the point of gaze fall within that region. There are two approaches:

a. Draw each AOI ‘by hand’ for each frame. This is can be reasonable for a small number of AOI and a fairly stable scene. The markers for the AOI can be dragged and re-sized in order to consistently overlay the actual regions in the scene as the video plays. It is potentially more time consuming than option 1, but is about as precise as mobile eye data analysis can get.

b. Use a form of image recognition software and have the software identify AOI. This approach works well if there is some contrast within the scene and you have primarily only one scene to deal. Here you create a video of the scene and identify the AOI for the software. It, in turn, uses this as a key to automatically identify the AOI in the recorded data files. One the AOI are identified, the process is the same as in (a) above. This potentially is a very quick process and results in data that is potentially very precise. It is almost too good to be true.

So that is where we stand at User First with our understanding of mobile tracking. We welcome corrections, other perspectives, elaboration, and the sharing of experiences as it pertains to this topic as we are ourselves learning each day about what is available, what is in the works, and how we might meet client needs in the future.

As we move forward with mobile tracking and expand on our experiences with different analysis methods in particular, we will keep you posted