Outlier Detection for Pedestrian Movement

Fast development of tracking devices has made trajectory outlier detection (TOD) possible and meaningful. Given a set of trajectories T , a TOD algorithm outputs a subset of T , of which trajectories are different from most of the other trajectories in some aspect(s). These trajectories, namely outliers, can indicate important or interesting information and are thus worth noticing. TOD techniques can be used for surveillance security, accident discovery, and many other purposes. Many of the existing TOD algorithms consider only spatial trajectory outliers. They can detect trajectories that follow an abnormal route or direction. While some existing algorithms are capable of detecting outliers in temporal aspects, like trajectories with abnormal time duration or speed, they have their own weaknesses. For example, they can be computationally expensive, or fail to detect important types of outliers. In this work, we aim to overcome these shortcomings of previous TOD algorithms. A novel grid-based TOD algorithm is proposed that is capable of detecting temporal-spatial outliers including density, direction, duration, and speed outliers with accuracy as well as fast calculation. The algorithm performs the following three main steps: (i) it calculates density, direction, duration, and speed features of all trajectories in the input set T ; (ii) it transforms feature information of trajectories into grid information; (iii) it examines each trajectory grid cell by grid cell. Following these three steps, outlying trajectories are extracted. By conducting experiments on several data sets including both simulated and real ones, the algorithm is shown to be efficient in detecting density, direction, duration, and speed outliers. It outperforms state-of-the-art TOD algorithms in various aspects.