Improved STNNet: A Benchmark from Detection, Tracking and Counting Crowds using Drones
Srikari Rallabandi, V Madhan, Anil Telaprolu, M. Sameena Nazeer · 2024
In the realm of computer vision, multi-object tracking in crowded scenes stands as a pivotal challenge with diverse applications, from surveillance systems to autonomous vehicles. Existing tracking methods, such as video surveillance for crowd control and public safety, often face difficulties in precisely associating noisy object detections and maintaining consistent labels across frames. To address these challenges, this paper presents ’Improved STNNet,’ an advanced framework for online Multi-Object Tracking (MOT). Building upon the foundational STNNet architecture, our improved model refines the decision-making process by incorporating deep reinforcement learning techniques. By formulating the online MOT problem as a Markov Decision Process (MDP), Improved STNNet learns a sophisticated policy for data association. Notably, the framework adeptly handles complexities such as object birth/death and appearance/disappearance, treating them as state transitions within the MDP. Through rigorous experimentation on benchmark datasets, including the MOT Challenge, our proposed Improved STNNet demonstrates superior performance, outperforming existing methods in challenging, crowded scenarios. The study not only provides compelling evidence of the efficacy of our approach but also opens avenues for advancing real-time video analysis applications, especially in dynamic, crowded environments. Some recent efforts have been devoted to constructing datasets for crowd-counting. For density map estimation, we used the dataset constructed and provided by STNNET, upon which this paper is extensively based.