The article proposes a new method for solving the problem of visual multiple object tracking (MOT) using a stochastic optimization approach. We called this method a Monte-Carlo Trajectory Optimization (MCTO).
The proposed method has two essential features, which, according to the authors, ensure the novelty of the proposed solution.
For most of the MOT algorithms currently used to combine detections into an object trajectory, initially for each detection (usually represented as a rectangular bounding box), an assumed connection between the detection and some object is built. Next, the parameters of the objects are estimated based on this association.
At the same time, in the case when the density of objects within the scene is large, the optimized list of possible connections can grow exponentially when several consecutive frames are included to the analysis.
We propose a different approach.
Initially, the search takes place in the space of possible object trajectories, which is well applicable for objects with a small number of degrees of freedom. From this space, a solution is selected by stochastic optimization using the observation model.
The second feature of the method is the use of a sliding window for optimization. This allows the method to work in a real time with some fixed delay in the result.
A set of experiments were carried out using the open datasets MOT17 and MOT20 to estimate the quality of tracking depending on used parameter values of the proposed method. The results obtained indicate the possibility of controlling the sensitivity of the method in a wide range.
A weak dependency of the resulting metrics on the number of optimization iterations is observed. The experiments have shown that the algorithm running time is linear to the window size and quadratic to the object density.
The proposed method was successfully applied to solve the problem of monitoring (determining the expected service time) queues at Moscow Sheremetyevo Airport.