Intelligent Camera Selection Decisions for Target Tracking in a Camera Network

Anil Sharma, Saket Anand, Sanjit K KaulWACV 2022

TL;DR

This paper tackles the challenge of efficiently tracking targets across a network of cameras by intelligently deciding which camera to query next. They propose a novel approach using reinforcement learning and a learned state representation (via an LSTM autoencoder) to make these decisions, significantly reducing unnecessary queries and improving tracking performance, even on large-scale networks.

What Problem Does This Paper Address?

Tracking a single object across multiple cameras in a network is a complex task, especially when cameras don't overlap. Existing methods often rely on making many re-identification (Re-ID) queries to find the target, which is computationally expensive and prone to false alarms. Predicting the next camera is also challenging because the time a target takes to transition between cameras (transition time) is often unknown and can vary. This uncertainty leads to difficulties in associating tracklets from different cameras and forming a complete trajectory. The paper aims to address this by developing an "intelligent" system that learns to make optimal camera selection decisions, minimizing redundant searches and improving overall tracking accuracy.

Methods

The core of the proposed method is a reinforcement learning (RL) framework, specifically using a Deep Q-Network (DQN). The system formulates camera selection as a Markov Decision Process (MDP).

  • State Representation: Instead of relying on hand-crafted state variables, the authors introduce a learned state representation. They use an LSTM-based autoencoder to encode the history of past camera selection decisions into a fixed-length latent vector. This learned representation is argued to be more effective and scalable, especially for larger camera networks.
  • Action Space: The agent can choose to query a specific camera or select a "null camera" action (Cx) to indicate no Re-ID query is needed.
  • Reward Function: A key contribution is a reward function designed to reduce the need for fully supervised training. It provides positive rewards for correctly identifying the target, a smaller reward for the "null camera" action, and a penalty for incorrect actions. This allows for semi-supervised training by providing rewards only at certain intervals.
  • DQN Architecture: A standard DQN architecture is employed, taking the learned state representation and other state variables (like target location and motion deltas) as input to predict Q-values for each action. The policy then selects the action with the highest Q-value.

Datasets

The proposed method was evaluated on four real-world datasets:

  • NLPR_MCT dataset: Used for camera selection and tracking.
  • DukeMTMC dataset: A large-scale dataset used for evaluation.
  • CityFlow dataset: Specifically scenarios S04 and S05, known for their large number of cameras.
  • WNMF dataset: Used for evaluating trajectory forecasting.

The paper notes that training and testing splits were aligned with existing state-of-the-art methods for fair comparison.

Key Insights and Ideas

  • Learned State Representation is Crucial: Hand-crafted state variables are insufficient for scaling to larger camera networks. An LSTM autoencoder effectively captures the history of camera selections, leading to better generalization and performance.
  • Semi-Supervised Training is Viable: The proposed reward function significantly reduces the annotation burden, allowing for effective training with limited supervision while achieving comparable performance to fully supervised methods.
  • Intelligent Queries Reduce Cost: By learning to predict when and where to query, the system drastically reduces the number of Re-ID calls, leading to faster processing and fewer false alarms.
  • Target-Agnostic Nature: The learned policy is shown to be effective across different target types and datasets, highlighting its generalizability.

Key Results

  • Camera Selection Performance: The proposed method (Ours and Ours-SS) achieves state-of-the-art or highly competitive results in terms of Precision (P) and F1-score (F1) across multiple datasets, often outperforming methods like nSteps [34] and CamSel [33]. For example, on NLPR-Set4, the method shows an absolute improvement of 39% over nSteps in Percentage Camera Handover (PCH).
  • Reduced Re-ID Queries: Figure 3 demonstrates that the proposed method makes significantly fewer Re-ID queries compared to other methods on the DukeMTMC dataset.
  • Tracking Performance: In multi-camera tracking (MCTA metric), the method shows strong performance, particularly on larger datasets, outperforming baseline methods like Neighbor.
  • Training Efficiency: Using the learned state representation (AE) leads to faster training times compared to training without AE.
  • Trajectory Forecasting: The method also shows promise in multi-camera trajectory forecasting, achieving better top-1 accuracy for next camera prediction and a significantly higher accuracy for camera selection compared to baselines.

Limitations and Future Work

  • Single Target Tracking: The current framework is designed for single-target tracking. To handle multiple targets, multiple parallel pipelines are run, which could be computationally intensive.
  • Dataset Size for AE Training: While the AE generalizes well, it's noted that the number of cameras at test time should be smaller than at training time.
  • Future Work: The paper suggests extending the framework to handle multiple targets more efficiently and exploring its application in other areas requiring intelligent decision-making in dynamic environments.

Summary updated

BibTeX
@inproceedings{sharma2022intelligent,
  author = {Anil Sharma and Saket Anand and Sanjit K Kaul},
  title = {{Intelligent Camera Selection Decisions for Target Tracking in a Camera Network}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
}
Figure