TL;DR
This paper tackles the critical safety issue of ensuring construction workers wear helmets by developing a real-time detection and tracking system. They combine an improved YOLOv5 object detector with the StrongSORT tracker, achieving high accuracy and robust tracking even in challenging scenarios like occlusion.
What Problem Does This Paper Address?
Construction site safety is paramount, and head injuries are a significant concern, often exacerbated by a lack of helmet usage. While object detection and tracking are well-established computer vision tasks, real-time monitoring of specific safety compliance like helmet wearing in dynamic and complex construction environments presents unique challenges. These include dealing with occlusions, rapid object motion, varying scales, and the need for high accuracy and speed to be practically useful for site management. Existing methods might struggle with the robustness required for such demanding applications.
Methods
The core of this work is a two-pronged approach:
- Object Detection: The authors leverage the YOLOv5s architecture, a lightweight and fast one-stage object detector. They introduce a key modification by replacing the standard CIOU (Complete Intersection over Union) loss function for bounding box regression with Focal-EIOU. This aims to improve convergence and give more weight to informative training samples, particularly for the challenging task of detecting helmets.
- Object Tracking: To maintain the identity of workers and their helmet status over time, they employ the StrongSORT algorithm. StrongSORT is presented as an improvement over DeepSORT, addressing issues like ID switching and loss during occlusions by incorporating more robust association strategies and potentially better feature extraction.
Datasets
The primary dataset used is the Safety Helmet Wearing Dataset, which contains over 8000 images. This dataset includes various construction scenarios, featuring workers both with and without helmets, as well as challenging conditions like dark scenes, occlusion, and complex task overlaps. The dataset was processed into three categories: target helmets, samples not wearing helmets, and heads as negative samples.
Key Insights and Ideas
- Focal-EIOU for Robust Bounding Box Regression: The choice of Focal-EIOU over CIOU for bounding box regression proved effective. This suggests that for tasks with specific object characteristics (like helmets), tailoring the loss function to better handle difficult cases and prioritize informative samples can yield significant improvements in detection accuracy.
- StrongSORT's Advantage in Occlusion: The paper highlights StrongSORT's superior performance in maintaining target IDs during occlusions compared to DeepSORT. This is crucial for real-time safety monitoring, as brief occlusions are common on construction sites, and losing track of a worker's helmet status can lead to false alarms or missed violations.
- YOLOv5s as a Lightweight and Effective Detector: The selection of YOLOv5s demonstrates that high performance doesn't necessarily require extremely large and computationally expensive models, making the system more deployable in real-world scenarios.
Key Results
- Improved Detection Accuracy: The YOLOv5s model with Focal-EIOU achieved a [email protected] of 95.4% and a helmet-wearing precision of 96.5% on the validation dataset. This represents an improvement of 0.3% and 0.8% respectively, over the original YOLOv5s with CIOU.
- High Inference Speed: The YOLOv5s detector achieved an inference speed of 100 images per second (FPS), with a model weight of only 14.4 MB.
- Robust Tracking: StrongSORT demonstrated a processing speed of 26.5 ms per frame for tracking. Crucially, it showed no ID switching or loss even when targets were occluded for extended periods, a significant advantage over DeepSORT which exhibited frequent ID switching.
- Overall System Performance: The combined YOLOv5s + StrongSORT system achieved a processing speed of 37 FPS.
Limitations and Future Work
- Limitations: The paper notes that in some instances of severe occlusion, a small number of objects could still be missed by the YOLOv5s detector. While StrongSORT is robust, the fundamental limitations of the detector can still impact the overall system.
- Future Work: The authors express interest in exploring how to apply this helmet-wearing tracking detection system to embedded terminal applications, suggesting a focus on further optimization for edge devices and real-time deployment in more diverse environments.
Summary updated