Abstract:The Long Short-Term Memory (LSTM) neural network based data association algorithm named as DeepDA for multi-target tracking in clutters is proposed to deal with the NP-hard combinatorial optimization problem in this paper. Different from the classical data association methods involving complex models and accurate prior knowledge on clutter density, filter covariance or associated gating etc, data-driven deep learning methods have been extensively researched for this topic. Firstly, data association mathematical problem for multitarget tracking on unknown target number, missed detection and clutter, which is beyond one-to-one mapping between observations and targets is redefined formally. Subsequently, an LSTM network is designed to learn the measurement-to-track association probability from radar noisy measurements and exist tracks. Moreover, an LSTM-based data-driven deep neural network after a supervised training through the BPTT and RMSprop optimization method can get the association probability directly. Experimental results on simulated data show a significant performance on association ratio, target ID switching and time-consuming for tracking multiple targets even they are crossing each other in the complicated clutter environment.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the data association problem in multi - target tracking in a clutter environment. Specifically, the paper focuses on how to effectively solve the data association, which is an NP - hard combinatorial optimization problem, in the presence of an unknown number of targets, missed detections and clutter. Traditional methods usually require complex models and accurate prior knowledge, such as clutter density, filter covariance or association thresholds, etc. In this paper, a deep data association network (DeepDA) based on long - short - term memory network (LSTM) is proposed. Through a data - driven method, it learns the association probabilities between measurements and targets, so as to directly obtain the association probabilities from radar noise measurements and existing trajectories. This method not only simplifies the model, but also improves the multi - target tracking performance in a complex clutter environment, especially in the case of target crossovers.
### Specific problems solved by the paper:
1. **Data association problem in multi - target tracking**: In multi - target tracking, data association is a crucial step, which involves determining the correspondence between sensor measurement values and targets. Traditional data association methods often require complex models and accurate prior knowledge, which are often difficult to obtain in practical applications.
2. **NP - hard combinatorial optimization problem**: The data association problem is essentially an NP - hard combinatorial optimization problem. Especially in the presence of a large amount of clutter and missed detections, traditional optimization methods (such as the Hungarian algorithm) may not be able to handle it effectively.
3. **Multi - target tracking in a complex environment**: In practical applications, the measurements received by radar or lidar may contain a large amount of clutter and missed detections, which makes multi - target tracking more difficult. The DeepDA method proposed in the paper aims to improve the multi - target tracking performance in such a complex environment.
### Main contributions of the paper:
1. **Extension of LSTM network**: The paper extends the LSTM network so that it can handle data association problems with 1 - 1, 1 - 0 and 0 - 1 constraints, meeting the data association requirements in the presence of clutter and missed detections.
2. **Supervised learning framework**: A supervised sequence - to - sequence learning framework without any prior knowledge is proposed, which transforms the integer programming problem into an association probability estimation problem.
3. **Experimental verification**: The effectiveness of the DeepDA method in a clutter environment is verified through simulation experiments, especially showing significant advantages in the case of target crossovers.
### Formula presentation:
- **Euclidean distance**:
\[
C_{ij}=\|Z_i - X_j\|^2
\]
where \(Z_i\) is the \(i\) - th measurement value and \(X_j\) is the state prediction of the \(j\) - th target.
- **0 - 1 decision variable**:
\[
x_{ij}=\begin{cases}
1 & \text{if } Z_i \text{ is assigned to } X_j\\
0 & \text{otherwise}
\end{cases}
\]
- **Loss function**:
\[
\mathcal{L}=\frac{1}{N}\sum_{i = 1}^{N}\left\|\beta_i-\hat{\beta}_i\right\|^2
\]
where \(\beta_i\) is the predicted association probability and \(\hat{\beta}_i\) is the true target - measurement association probability.
### Summary:
By introducing the deep data association network (DeepDA) based on LSTM, the paper successfully solves the data association problem in multi - target tracking, especially showing significant advantages in the presence of a large amount of clutter and missed detections. This method not only simplifies the model, but also improves the accuracy and robustness of multi - target tracking.