“…Approaches [2,4,8,19] considering both the past and the future information typically require low-level observations such as foreground, tracklet, or trajectory, etc. These types of low level observations can be obtained by background modelling [35] (to acquire foreground), or by associating confident responses of a human detector, head detector or part-based detector into tracklets [7,12,17,25,33,43,44,45] (this is the most popular approach since significant progress has been made in the detection field [13,16]), or by estimating trajectories based on the KLT tracker [36] or Kalman Filter [12]. Then, these types of low level observations are associated by optimisation methods, such as Markov Chain Monte Carlo (MCMC) [35], Dynamic Programming, Hungarian algorithm [33,43], greedy bipartite algorithm [34], network flow [41] and K-Shortest Paths (KSP) algorithm [5].…”