Classification of high-level behaviors in video data has been identified by the Federal Highway Administration (FHWA) as a gap in development of practical automated image processing to enhance the value of large research datasets (e.g., SHRP2) and other applications. The researchers will take several approaches to practical automated identification of high-level behaviors. For this problem, they define behavioral semantics, or primitives, that they assume are identifiable by existing algorithms, including those developed by team members. Semantics are defined by top-down analysis based on human judgment as well as bottom-up analysis from image processing. The link between semantics and high-level behaviors is found using: 1) machine learning; 2) propagation of movement using intent-based cost models; and 3) statistical relationships. These approaches may be most suited to different high-level behavior identification, and the researchers will develop them in parallel. To ensure practical implementation, the researchers also will focus some effort on speed and accuracy. Their approach to these is general purpose and uses context cues to identify which parts of the image to process and which images to process, and to provide priors that improve classification performance.