Abstract
Data for the association for a multi-target tracking (data association), proposed a network-based method of optimizing flow (network flow) of. The maximum a posteriori probability (maximum-a-posteriori: MAP ) mapping data association problem to satisfy the non-overlapping tracks (non-overlap) flow network cost constraints (cost-flow network). By minimizing cost network flow algorithm (a min-cost flow algorithm) , you may find an optimal association data. The network expansion: shutter contain an explicit model (EOM), can be used to process the presence of long (long-term) inter-object occlusion (inter-object occlusions) tracking phenomenon. On the basis of the original algorithm, which can be solved by iteration EOM-based network. Initialization and termination trajectory and potential false observation (potential false observations) by the inner frame modeled. The method of high efficiency, and does not require assuming prune (hypotheses pruning). Disclosed in the two data sets pedestrians, by comparison with the performance results of previous studies, the effect of improving the description of the method.
1. Introduction
Robustness target detection and tracking are very important for many computer vision tasks. The method of our discussion are: the detection result of each target frame image as input and associates the detection result to look up the target track. Not all targets can be detected in each frame, and there may be error detection, there are some targets may be obscured by other goals; these factors make the data associated with becoming a difficult task.
Some methods, such as [1,2], attempt to resolve ambiguities in each frame. Others, such as [9, 10] is to use more global information. However, these alternatives search space with the increase in the number of frames of exponential growth, which requires strictly limit the search window and assumptions pruning. They also usually assume that all test results are correct, but in fact does not guarantee always correct.
We propose an efficient global data association method, compared to the previous method, it can find the optimal solution for the many long sequences of (window). In our method, the data is associated with a group defined for a given target detection result as MAP estimation input observations. The model assumed a non-overlapping track is not cost-phase alternating current flow path in the network (disjoint flow path); observation likelihood and the transition probability (observation likelihood and transition probabilities) is modeled as a stream costs. With a minimum cost flow algorithm to obtain the global optimum track association. Tracking process as in the case of long blocked by adding blocking constraint nodes and the network (considering only inter-object occlusions occlusion between the target), the establishment of the explicit occlusion model (EOM). In the original minimum cost flow algorithm, solved using minimum cost flow (a minimal cost flow) based network EOM iterative method. Trace initialization, termination and inference object occlusion intrinsic behavior of this method, it can be inferred from the correlation result information. Examples of a blocked and undetected inference from the tracking results are shown in Fig.
The remainder of this paper is organized as follows. Related work is discussed in Section 2. Section 3 describes the MAP formulation and its global optimal solution. Section 4 describes explicit occlusion model (EOM) and an iterative solution method. Section 5 given implementation details. The results see section 6. Conclusion Section 7.
2. Related work
To track a plurality of targets, a process frame by frame (or within a small window of time) to make decisions related data, such as [1,2]. While this approach has shown good performance, but consider more frames before making a decision is usually associated with it should help to better overcome the ambiguity and long obscured or missed due to false detection.
Many more information on using global methods have been tried to overcome detect errors. One strategy is to optimize the entire sequence of a trajectory through time; this has been adopted in the process based on dynamic programming, as described [5,6]. Then use a combination of these greedy strategy to track and deal with potential conflicts. Since a single track separately optimize these methods it is difficult to model occlusion. Another method is to optimize a plurality of tracks; multiple hypothesis tracking (MHT: multi-Hypothesis Tracking) [3] and the joint probability data association filter (JPDAF: Joint Probabilistic Data Association Filters ) [4] are two representative example of. Further, in [10], the trajectory estimation and hypothesis detection combined with Boolean quadratic programming. Since it is assumed to be a combination of the search space (Combinatorial), such methods can only be optimized for a limited time window must also be assumed pruning. Sampling methods such as MCMC [9] is also used to solve similar problems. In these methods, the shutter is usually modeled as merging and splitting the trajectory (merging and splitting of trajectories).
Tracklet Stitching [8] based on linear programming (LP) tracking [7] The other two methods is sought over the entire sequence optimization of all tracks simultaneously. [8] First generation tracklets, which is formed of a detection response (detection responses) conservative packet track segment (fragments of tracks). Hungary partitioning algorithm then connected tracklets. This method assumes that all tracklets correspond to the real target track, it is difficult to apply to the original detection result , i.e., a number of possible false alarm (false alarms) each frame. [7] Construction of a sub-set of graphs for each object track, the edges between the sub-FIG (Edges) represents the interaction objects (object interactions). Then solving the multi-path search sub-linear programming problem by Atlas and rounding approximately. It is assumed that between the target position needs (inter-object positions) is relatively stable, the number of fixed target.
3. Our approach
We define data related to MAP problem. Then map MAP network flow problem of the cost (a cost-flow network), and is solved by the minimum cost flow algorithm (a min-cost flow algorithm) . Such mapping is based on looking to find non-overlapping target locus and similarity exists between the edge disjoint paths (edge-disjoint paths) in the drawing ; the latter can be effectively solved by the network flow algorithm. We first proposed the idea, and then provide the minimum cost flow Solution.
3.1 MAP in non-overlapping constraint
3.2 minimum cost flow Solution
One example of FIG. 2 flow network cost, comprising three time steps 9 and observations
• Construct the graph G(V, E, C, f) from observation set X
• Start with empty flow
• WHILE ( f(G) can be augmented )
– Augment f(G) by one.
– Find the min cost flow by the algorithm of [12].
– IF ( current min cost < global optimal cost )
Store current min-cost assignment as global optimum.
• Return the global optimal flow as the best association hypothesis
Algorithm 1:Find MAP trajectories by min-cost flow
算法1为数据关联提供了一个通用框架。该方法不同于单独优化各轨迹的方法，也不受困于假设空间组合激增的方法，它能够有效率地找到全局最优解。接下来，我们扩展我们的方法来处理长时间遮挡的跟踪。