Behavior Analysis Using a Multilevel Motion Pattern Learning Framework

Abstract

The increasing availability of video data, through existing traffic cameras or dedicated field data collection, and the development of computer vision techniques pave the way for the collection of massive data sets about the microscopic behavior of road users. Analysis of such data sets helps in understanding normal road user behavior and can be used for realistic prediction of motion and computation of surrogate safety indicators. A multilevel motion pattern learning framework was developed to enable automated scene interpretation, anomalous behavior detection, and surrogate safety analysis. First, points of interest (POIs) were learned on the basis of the Gaussian mixture model and the expectation maximization algorithm and then used to form activity paths (APs). Second, motion patterns, represented by trajectory prototypes, were learned from road users' trajectories in each AP by using a two-stage trajectory clustering method based on spatial then temporal (speed) information. Finally, motion prediction relied on matching at each instant partial trajectories to the learned prototypes to evaluate potential for collision by using computing indicators. An intersection case study demonstrates the framework's ability in many ways: it helps reduce the computation cost up to 90%; it cleans the trajectory data set from tracking outliers; it uses actual trajectories as prototypes without any pre- and postprocessing; and it predicts future motion realistically to compute surrogate safety indicators.

Analysing cyclist behaviour at cycling facility discontinuities using video data

“…Another challenge of clustering trajectories is the representation of each cluster (motion pattern): contrary to methods like k-means applied to fixed-length vectors, trajectories cannot be easily averaged. An original model and clustering algorithm based on the LCSS were proposed in (Saunier, Sayed, & Lim, 2007) and refined in (Mohamed & Saunier, 2015) where each cluster is represented by an actual trajectory and trajectories are assigned to the motion pattern cluster based on their highest similarity. Motion pattern learning was initially developed and applied to motion prediction to compute surrogate measures of safety like time to collision (Mohamed & Saunier, 2015;Saunier et al, 2007).…”

“…The final step in the methodology (step 5) relies on the cluster model and clustering algorithm developed in previous work (Mohamed & Saunier, 2015;Saunier & Sayed, 2006). Each cluster is represented by an actual road user trajectory (the longest).…”

“…road user trajectories at a fine temporal scale. To that end, our study applies computer vision techniques and trajectory clustering methods to video data, as done in a few past road user behaviour and safety studies (Laureshyn, Goede, Saunier, & Fyhri, 2017;Mohamed & Saunier, 2015;Saunier & Sayed, 2006). This study makes use of video data to compare cyclist behaviour between sites with and without a discontinuity by automatically extracting road user trajectories from video data.…”

Driver Behavior Classification at Stop-Controlled Intersections Using Video-Based Trajectory Data

“…State-of-the-art machine learning techniques have great potential to solve complex problems, including understanding complex road user behavior [31][32][33][34][35][36]. Different studies have investigated the pattern of driver behavior using the machine learning approach [32,37,38].…”

“…State-of-the-art machine learning techniques have great potential to solve complex problems, including understanding complex road user behavior [31][32][33][34][35][36]. Different studies have investigated the pattern of driver behavior using the machine learning approach [32,37,38]. For example, Mohamed and Saunier [32] introduced a multi-level motion pattern learning framework for understanding driver behavior in an unsupervised pattern recognition approach.…”