YOLO-PL: Helmet wearing detection algorithm based on improved YOLOv4
Haibin Li,Dengchao Wu,Wenming Zhang,Cunjun Xiao
DOI: https://doi.org/10.1016/j.dsp.2023.104283
IF: 2.92
2024-01-01
Digital Signal Processing
Abstract:Workplace safety accidents are a pervasive issue worldwide. According to the National Work Safety Supervision Administration, a striking 67.95 % of construction accidents occur due to workers not wearing helmets. Existing helmet-wearing detection algorithms, however, tend to underperform in real-world scenarios where challenges such as smaller helmet areas in images, complex backgrounds, and object occlusions are present. Additionally, these models have a considerable amount of parameters, which impedes their practical deployment. This study proposes a novel, lightweight helmet detection algorithm, YOLO-PL, based on YOLOv4, to address these challenges. Initially, we designed the YOLO-P algorithms. YOLO-P algorithms optimize the network structure by refining its ability to detect small objects and improving the anchor assignment in the detection head. We design the Enhanced PAN (E-PAN) structure to merge the higher-layer, low-noise information with the lower-layer information based on the Path Aggregation Network (PAN). The YOLO-P algorithm improves detection accuracy by using the E-PAN structure. Subsequently, while preserving the performance of the YOLO-P algorithm, we enhanced its design for lightness. We proposed the Dilated Convolution Cross Stage Partial with X res units (DCSPX) module based on the Cross Stage Partial (CSP) structure, replacing the Spatial Pyramid Pooling (SPP) module with it. Additionally, we designed a Lightweight VoVNet (L-VoVN) structure based on the architecture of VoVNet, introduced a lightweight Max-Pooling (MP) down-sampling method, and fine-tuned the Swish activation function, which led to the final YOLO-PL algorithm. YOLO-PL significantly reduces the parameters in YOLO-P, thus achieving state-of-the-art performance that surpasses current object detectors like YOLOv5 and v7 in safety helmet detection. Moreover, our model exhibits substantial improvements in robustness and deployability, demonstrating considerable potential for practical implementations in industry.
engineering, electrical & electronic