Observations of the dynamics of wildland fires are limited, particularly at the fine temporal and spatial scales at which fireline dynamics occurs. Moreover, a good understanding of the fire behavior requires a clear picture of the three-dimensional winds in the fire, including the flaming combustion zone and the convective motions produced by the fire. These motions determine the convergence pattern superimposed upon the surface winds that in turn direct and propagate the fire, and in more extreme crown fire behavior, are an intrinsic part of the dynamics of fire propagation. To address this problem, an image flow analysis technique was used to analyze sequences of radiant temperature data (from the 3.4-5.0 micron range from an Inframetrics infrared imager) and determine atmospheric motions in and around the flaming combustion zone and convective column produced by fires. This work summarizes the observations and analysis of data collected in both crown fires during the FROSTFIRE experiment and Colorado wildfires and grass fires during prescribed burns. In crown fires during FROSTFIRE, the analysis reveals a picture that is dynamically complex, and defies the simple notion of one large convective plume or many tree-scale plumes that rise separately, simply accelerating under the force of buoyancy. Instead, the picture that emerges is that of a sequence of upslope surges of many convective plumes, representing a scale larger than individual trees, that shoot some distance uphill near to the ground parallel to the surface before rising vertically. The analysis shows that velocity components parallel to the surface of 10 m/s are routine, with frequent bursts of 25 m/s, and peak surges of 30 m/s are detectable less than 15 m above the base of the detectable fire. Vertical velocity statistics suggest that maximum updrafts of over 60 m/s are common in this fire. The component of velocity parallel to the approximately 20 degree slope frequently had peaks of 35-40 m/s. We emphasize that the speed of extremely high temperature air near the surface is not based upon ambient wind speeds (which were weak), but a result of complex fire-atmosphere dynamics that transfer vertical momentum into surges along the surface. Fire spread rates in a flank of a fire exposed to our view were on average 0.75 m/s, with peaks of 1.26 m/s over 10-sec periods, despite the weak ambient winds. Other fire observations, which will be described, allowed us to detect and quantitatively analyze processes such as spotting.