Image convolution techniques integrated with YOLOv3 algorithm in motion object data filtering and detection

Mai Cheng, Mengyuan LiuScientific 2024 · DOI

Abstract

In order to address the challenges of identifying, detecting, and tracking moving objects in video surveillance, this paper emphasizes image-based dynamic entity detection. It delves into the complexities of numerous moving objects, dense targets, and intricate backgrounds. Leveraging the You Only Look Once (YOLOv3) algorithm framework, this paper proposes improvements in image segmentation and data filtering to address these challenges. These enhancements form a novel multi-object detection algorithm based on an improved YOLOv3 framework, specifically designed for video applications. Experimental validation demonstrates the feasibility of this algorithm, with success rates exceeding 60% for videos such as "jogging", "subway", "video 1", and "video 2". Notably, the detection success rates for "jogging" and "video 1" consistently surpass 80%, indicating outstanding detection performance. Although the accuracy slightly decreases for "Bolt" and "Walking2", success rates still hover around 70%. Comparative analysis with other algorithms reveals that this method's tracking accuracy surpasses that of particle filters, Discriminative Scale Space Tracker (DSST), and Scale Adaptive Multiple Features (SAMF) algorithms, with an accuracy of 0.822. This indicates superior overall performance in target tracking. Therefore, the improved YOLOv3-based multi-object detection and tracking algorithm demonstrates robust filtering and detection capabilities in noise-resistant experiments, making it highly suitable for various detection tasks in practical applications.

Introduction

Real-time video surveillance remains a significant challenge for standard deep learning frameworks. Issues such as dense target overlap, environmental noise (e.g., rain, fog, or low light), and the inherent trade-off between processing speed and detection accuracy often lead to missed detections or imprecise localization. This paper introduces an enhanced YOLOv3-based framework specifically optimized for these rigors, achieving a tracking accuracy of 0.822 and maintaining high precision even in high-noise environments.

Methodology

The researchers built upon the Darknet-53 backbone of YOLOv3, introducing three specific architectural and algorithmic improvements designed to refine how the network processes spatial information and training data.

Inception-Inspired Convolutional Blocks

To increase the receptive field without ballooning the parameter count, standard $5 \times 5$ kernels were replaced with a factorized combination of $1 \times 1$, $3 \times 1$, and $1 \times 3$ kernels. As shown in Figure 3, this "factorization" strategy reduces the number of parameters by nearly one-third while maintaining the network's ability to extract complex features through residual connections.

Factorized convolutional block with asymmetrical kernels and residual connections.

Improved K-Means with Data Filtering

Standard YOLOv3 uses K-means clustering on training datasets to determine optimal "anchor box" shapes. However, outliers can skew these results. The authors introduced a pre-processing "judgment module" (Figure 4) to filter out invalid coordinate data from annotation files. By calculating aspect ratios and coordinate differences, the module ensures that anchor boxes are more representative of actual targets, improving localization precision.

Flowchart of the improved judgment module for filtering coordinate data prior to K-means clustering.

Optimized Image Partitioning

The researchers identified that grid size significantly impacts the resolution of object centers. After testing $7 \times 7$, $10 \times 10$, and $14 \times 14$ grids, they determined that a $10 \times 10$ partition provides the optimal balance—offering a significant accuracy boost over smaller grids without the prohibitive computational cost of higher resolutions.

Experimental Results

The model was evaluated using industry-standard benchmarks including COCO, MOT16, and OTB-2015. Key performance metrics include:

  • Tracking Accuracy: The proposed algorithm achieved a tracking accuracy of 0.822, outperforming established methods like DSST (0.603) and MTS (0.639).
  • Noise Resistance: In extreme noise conditions (SNR of 0.1), where standard YOLOv3 accuracy dropped to 0%, the improved model maintained a precision of 77.78%.
  • Real-Time Performance: Utilizing the $10 \times 10$ partitioning scheme, the system achieved a frame rate of 43.24 FPS, making it highly suitable for live monitoring.

Conclusion

The improved YOLOv3 framework demonstrates robust filtering and detection capabilities, particularly in noise-resistant scenarios. By integrating factorized convolutional layers and a coordinate judgment module, the algorithm addresses common limitations like false positives and imprecise localization. Future work will focus on adapting these techniques to handle dynamic backgrounds, expanding the application to mobile robotics and autonomous vehicle safety systems.

Summary updated

Abstract

In order to address the challenges of identifying, detecting, and tracking moving objects in video surveillance, this paper emphasizes image-based dynamic entity detection. It delves into the complexities of numerous moving objects, dense targets, and intricate backgrounds. Leveraging the You Only Look Once (YOLOv3) algorithm framework, this paper proposes improvements in image segmentation and data filtering to address these challenges. These enhancements form a novel multi-object detection algorithm based on an improved YOLOv3 framework, specifically designed for video applications. Experimental validation demonstrates the feasibility of this algorithm, with success rates exceeding 60% for videos such as "jogging", "subway", "video 1", and "video 2". Notably, the detection success rates for "jogging" and "video 1" consistently surpass 80%, indicating outstanding detection performance. Although the accuracy slightly decreases for "Bolt" and "Walking2", success rates still hover around 70%. Comparative analysis with other algorithms reveals that this method's tracking accuracy surpasses that of particle filters, Discriminative Scale Space Tracker (DSST), and Scale Adaptive Multiple Features (SAMF) algorithms, with an accuracy of 0.822. This indicates superior overall performance in target tracking. Therefore, the improved YOLOv3-based multi-object detection and tracking algorithm demonstrates robust filtering and detection capabilities in noise-resistant experiments, making it highly suitable for various detection tasks in practical applications. It can address inherent limitations such as missed detections, false positives, and imprecise localization. These improvements significantly enhance the efficiency and accuracy of target detection, providing valuable insights for researchers in the field of object detection, tracking, and recognition in video surveillance.
BibTeX
@inproceedings{cheng2024image,
  author = {Mai Cheng and Mengyuan Liu},
  title = {{Image convolution techniques integrated with YOLOv3 algorithm in motion object data filtering and detection}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-57799-0},
}
Figure