Personnel localization method for low-visibility environments based on improved YOLOv3
-
Graphical Abstract
-
Abstract
In coal mines, inadequate lighting and dust obstruction result in personnel targets captured by video monitoring systems appearing as small or low-visibility objects in two-dimensional images. The original YOLOv3 network's Darknet53 feature pyramid structure was insufficient for effectively extracting and preserving detailed information about these targets, leading to inaccurate localization. To address this issue, personnel localization method for low-visibility environments based on improved YOLOv3 was. First, the clarity of coal mine monitoring videos under low-visibility conditions was enhanced using a combination of β function mapping and inter-frame information enhancement techniques. Next, Darknet53 in YOLOv3 was replaced with the lighter Darknet-19, and CIoU was introduced as the loss function to optimize personnel target identification in the enhanced video. Finally, the identified targets were projected from two-dimensional space to three-dimensional space based on the mapping model, completing the personnel localization process. Experiments conducted on monitoring videos from a coal mine in low-visibility conditions revealed the following findings: ① After applying the improved YOLOv3, the brightness, visibility, and various evaluation metrics (average gray level, average contrast, information entropy, and gray spectral bandwidth) of the video frames demonstrated significant improvements compared to the original videos. There was a substantial enhancement in overall lighting conditions and contrast, facilitating better differentiation between targets and backgrounds, thereby validating the effectiveness of the image enhancement techniques employed. ② The improved YOLOv3 accurately identified personnel in the video frames, with no instances of missed detections. ③ Using calibrated objects or manually annotated real three-dimensional positions as benchmarks, the deviation between the projected results and the actual positions was calculated (covering distance deviations in the X, Y, and Z directions). The deviations in both the X and Y directions were below 0.2 m, while the deviation in the Z direction was below 0.002 m, indicating a high mapping effect and localization accuracy of the constructed mapping model.
-
-