Object Counting By Tracking For Complex Video Sequences

Object counting within videos plays a crucial role in numerous applications, including crowd monitoring, traffic analysis, surveillance, and resource management. The ability to reliably count and track objects is critical for ensuring safety in public spaces, managing traffic flows, and analyzing behavior patterns in various domains. However, the task of counting objects in video frames poses significant challenges, including overlapping objects, occlusions, and repetitive counting across multiple frames. Previous methods such as counting-by-detection, regression, and segmentation have demonstrated limited effectiveness in these complex scenarios. Specifically, it was shown that counting-by-detection struggles with overlapping objects, while regression-based approaches discard crucial spatial information, and segmentation techniques depend heavily on image quality and struggle with occlusions.This paper addresses the abovementioned challenges and presents a new system encompassing advanced object detection, re-identification, and multi-object tracking algorithms, including the YOLOv10 model and ByteTrack. The proposed system performance was further enhanced by adding a custom Non-Maximum Suppression (NMS) function and a minimum frame count threshold (MIN_FRAMES) to reduce false positives. We conducted several experiments which showed improved accuracy and reliability of the tracking system, particularly in challenging environments with occlusions and multiple appearances, leading to 91.43% average improvement in RMSE when compared to the state-of-the-art Ultralytics object counting method and a 60.53% average improvement in RMSE compared to not adding the NMS and MIN FRAMES to our system.