2 Multi-Tracked Paths (MTP)
3 -------------------------
7 This is a very simple implementation of a variant of the k-shortest
8 paths algorithm (KSP) applied to multi-target tracking, as described
11 J. Berclaz, E. Turetken, F. Fleuret, and P. Fua. Multiple Object
12 Tracking using K-Shortest Paths Optimization. IEEE Transactions on
13 Pattern Analysis and Machine Intelligence (TPAMI), 33(9):1806-1819,
16 This implementation is not the reference implementation used for the
17 experiments presented in this article. It uses a Dijkstra with a
18 Binary Heap for the min-queue, and not the optimal Fibonacci heap.
20 This software package includes three commands:
22 - mtp is the generic command to use in practice. It takes tracking
23 parameters as input, and prints the tracked trajectories as
24 output. The format for these parameters is given at the bottom of
27 - mtp_example creates a tracking toy example, and runs the tracking
28 algorithm on it. It gives an example of how to use MTPTracker on a
29 configuration produced dynamically, and produce a test input file
32 - mtp_stress_test creates a larger problem with a lot of noise and
33 multiple trajectories, to check the behavior of the code under
34 slightly more complex situations.
38 This software should compile with any C++ compiler. Under a unix-like
39 environment, just execute
44 It will create a synthetic dummy example, save its description in
45 tracker.dat, and print the optimal detected trajectories.
49 ./mtp --verbose --trajectory-file result.trj --graph-file graph.dot tracker.dat
51 It will load the file tracker.dat saved by the previous command, run
52 the detection, save the detected trajectories in result.trj, and the
53 underlying graph with occupied edges in graph.dot.
55 If you do have the graphviz set of tools installed, you can produce a
56 pdf from the latter with the dot command:
58 dot < graph.dot -T pdf -o graph.pdf
62 The two main classes are MTPGraph and MTPTracker.
64 The MTPGraph class contains a directed acyclic graph (DAG), with a
65 length for each edge -- which can be negative -- and has methods to
66 compute the family of paths in this graph that globally minimizes the
69 If there are no path of negative length, this optimal family will be
70 empty, since the minimum total length you can achieve is zero. Note
71 that the procedure is similar to that of KSP, in the sense that the
72 family it computes eventually is globally optimal, even if the
73 computation is iterative.
75 The MTPTracker takes as input
77 (1) a spatial topology composed of
79 - a number of locations
81 - the allowed motions between them (a Boolean flag for each pair
84 - the entrances (a Boolean flag for each location and time step)
86 - the exits (a Boolean flag for each location and time step)
88 (2) a number of time steps
90 (3) a detection score for every location and time, which stands for
92 log( P(Y(l,t) = 1 | X) / P(Y(l,t) = 0 | X) )
94 where Y is the occupancy of location l at time t and X is the
95 available observation. Hence, this score is negative on locations
96 where the probability that the location is occupied is close to
97 0, and positive when it is close to 1.
99 From this parameters, an MTPTracker can compute the best set of
100 disjoint trajectories consistent with the defined topology, which
101 maximizes the overall detection score (i.e. the sum of the detection
102 scores of the nodes visited by the trajectories). In particular, if no
103 trajectory of total positive detection score exists, this optimal set
104 of trajectories is empty.
106 An MTPTracker is a wrapper around an MTPGraph. From the defined
107 spatial topology and number of time steps, it builds a graph with one
108 source, one sink, and two nodes per location and time. The edges from
109 the source or to the sink, or between these pairs of nodes, are of
110 length zero, and the edges between the two nodes of such a pair have
111 negative lengths, equal to the opposite of the corresponding detection
112 scores. This structure ensures that the trajectories computed by the
113 MTPTracker will be node-disjoint, since the trajectories computed by
114 the MTPGraph are edge-disjoint.
116 The file mtp_example.cc gives a very simple usage example of the
117 MTPTracker class by setting the tracker parameters dynamically, and
118 running the tracking.
120 The tracker data file for MTPTracker::read has the following format,
121 where L is the number of locations and T is the number of time steps:
123 ---------------------------- snip snip -------------------------------
126 bool:allowed_motion_from_1_to_1 ... bool:allowed_motion_from_1_to_L
128 bool:allowed_motion_from_L_to_1 ... bool:allowed_motion_from_L_to_L
130 bool:entrance_1_1 ... bool:entrance_1_L
132 bool:entrance_T_1 ... bool:entrance_T_L
134 bool:exit_1_1 ... bool:exit_1_L
136 bool:exit_T_1 ... bool:exit_T_L
138 float:detection_score_1_1 ... float:detection_score_1_L
140 float:detection_score_T_1 ... float:detection_score_T_L
141 ---------------------------- snip snip -------------------------------
143 The method MTPTracker::write_trajectories writes first the number of
144 trajectories, followed by one line per trajectory with the following
147 ---------------------------- snip snip -------------------------------
148 int:traj_number int:entrance_time int:duration float:score int:location_1 ... int:location_duration
149 ---------------------------- snip snip -------------------------------