Abstract
While deep learning has significantly advanced object detection, video-based analysis still faces challenges like "detection jitter"—the unstable flickering of object counts between frames—and inaccuracies in high-density or overlapping scenes. This paper introduces YOLOv3-ANV (Abnormal Number Version), a framework designed to stabilize multi-object detection in video streams. By integrating a "judge" method that switches between standard detection and pixel-based density estimation, and an optimization algorithm for temporal smoothing, the proposed method achieves a 5.4% average accuracy improvement over existing models on the Wanfang sports competition and NWU campus datasets.
Introduction
Modern convolutional neural networks (CNNs) like YOLO and Faster R-CNN excel at detecting objects in static images. However, when applied to video, slight variations in lighting, motion blur, or occlusion can cause the model to drop or add objects inconsistently across consecutive frames. This phenomenon, known as detection jitter, makes it difficult to accurately calculate pedestrian flow or monitor athlete behavior in real-time. Furthermore, standard bounding-box detectors often fail in crowded environments where objects are small and highly overlapping. This research addresses these issues by proposing a two-stage framework that ensures both temporal stability and counting accuracy in dense scenes.
Methodology
The YOLOv3-ANV framework employs a hybrid approach to handle varying scene densities and temporal inconsistencies.
1. The "Judge" Method (Hybrid Detection)
To optimize for different scenarios, the system uses a threshold-based switch (set at $N=25$).
- Low Density: When the detected object count is below the threshold, the system relies on YOLOv3, which provides high-precision bounding boxes for sparse scenes.
- High Density: If the count exceeds the threshold, the system switches to a pixel-based population density estimation algorithm. This path uses background subtraction to perform foreground segmentation, identifying "motion prospects."
Features such as perimeter and area are extracted from these segments and passed through a linear regression model to estimate the total count. This method is significantly more robust than bounding-box detection in highly crowded environments.
2. The Optimization Method (Jitter Prevention)
To eliminate temporal instability, the authors implemented a smoothing algorithm. The system defines an "error correction region" (a temporal window) equivalent to 1/3 of the video's frame rate. If a frame's detection count deviates significantly from its neighbors, the algorithm replaces that value with the maximum count found within the local window. This heuristic assumes that "leakage" (missing an object) is a more frequent error than false positives, effectively stabilizing the detection sequence.
Experimental Results
The proposed model was evaluated using the Wanfang Sports Competition Dataset (covering rugby, hockey, basketball, and horse racing) and a custom NWU Campus Dataset for pedestrian monitoring.
- Accuracy Improvements: YOLOv3-ANV achieved an average improvement of 5.4% over standard methods. In specific tasks like Rugby competition monitoring, accuracy reached 80.71%, compared to 71.57% for standard YOLOv3.
- Pedestrian Monitoring: On the NWU campus dataset, the model reached a high accuracy of 90.11%, demonstrating its effectiveness for security and flow analysis.
- Stability: The optimization algorithm successfully smoothed the output sequences, providing a reliable basis for "abnormal number" alerts in security contexts.
Conclusion
YOLOv3-ANV provides a robust solution for real-time multi-object detection in video. By combining the spatial precision of YOLOv3 with the crowd-resilience of pixel-based density estimation and a temporal smoothing layer, the framework effectively mitigates detection jitter. Future work will focus on expanding training data diversity and optimizing the model for lower-power edge devices.
Summary updated