Abstract:In [11], we showed that an efficient DP-TBD implementation for MTT can be achieved by using the following strategies: 1. Exploiting the structure of the MTT problem: the high-dimensional maximization in DP-TBD can be solved suboptimally in a cheaper but accurate way by using the inherent structure of the tracking model [6]. When targets are clustered into well-separated groups, there is no measurement ambiguity between distant targets; therefore, the target groups can be treated independently. This idea has already been used in particle filter-based MTT methods [6-8] to enhance the sampling efficiency, but it is an unexplored area for DP-TBD. In [11], we showed that, by incorporating this idea into DP-TBD, the high-dimensional maximization can be broken down to several maximizations of lower dimension; thus, the computational cost is significantly reduced. Targets in the same cluster are still considered jointly, so the tracking performance is resilient to target interference when targets are in close proximity. 2. Generalized detection procedure: existing DP-TBD algorithms all use the same detection procedure: after DP integration, the single-target or multitarget state with the global maximum merit is selected and tested against a detection threshold. If the global maximum merit exceeds the threshold, the estimated trajectory is declared and recovered; otherwise the null hypothesis of no target is accepted. In [11], motivated by our analysis of the single-target DP merit function, a generalized detection procedure is developed. A key advantage of this detection procedure is that it requires only a single-target state DP integration without regard for how many targets exist. After the single-target DP integration, not only the global maximum but also every merit is tested against the detection threshold. All the exceeded elements are retained to form an estimation candidate group from which the partitions of the estimated multitarget state, i.e., individual target states, will be extracted recursively in the subsequent processing. The estimated number of targets is then the number of the extracted individual target states. As a result, the dimension of the multitarget state no longer needs to be predetermined during the DP implementation as in existing algorithms. Single-target and multitarget scenarios are handled within one framework instead of using a multiple hypothesis algorithm.

Multi-target video tracking algorithm based on Camshift and TLD framework

Robust Visual Tracking Via CAMShift and Structural Local Sparse Appearance Model

APPTracker: Improving Tracking Multiple Objects in Low-Frame-Rate Videos

APPTracker Plus : Displacement Uncertainty for Occlusion Handling in Low-Frame-Rate Multiple Object Tracking

Detecting and Tracking Dynamic Objects in Complex Environments

Multi-target Tracking by Learning Local-to-global Trajectory Models

Camshift based real-time multiple faces match tracking

Multi-target tracking via dynamic-programming based track-before-detect

Research on Video Object Tracking Based on Improved Camshift Algorithm

An Efficient Multi-Frame Track-Before-Detect Algorithm for Multi-Target Tracking

Target Tracking Based on Camshift and Kalman Filter Algorithm

A Tracking Approach for Low Observable Target Using Plot-Sequences of Multi-Frame Detection

Student Highlight: Dynamic Programming-Based Track-Before-detect Approach to Multitarget Tracking

Pedestrian detection and tracking based on TLD framework

Video Object Tracking Based on SSD and Camshift

Learning to track multiple targets

Research on Target Tracking Based on TLD Algorithm

Long-term tracking algorithm using deep features and a single shot multibox detector.

Object Tracking Algorithm with Sparse Prototype Based on TLD

LGTrack: Exploiting Local and Global Properties for Robust Visual Tracking

Tracking of infrared small-target based on improved Mean-Shift algorithm