Hardhat Detection For Factory Surveillance Videos

Hardhat detection is an important part of safety work.With the increasing number of surveillance equipment and surveillance videos,it has become an urgent need to effectively utilize these surveillance video resources in industrial environments.In recent years,deep learning methods have made rapid progress in object detection and image classification,but most of them are based on public datasets.There are still many problems in accuracy and speed when applied in industrial scenarios.We use pedestrian detection and hardhat detection to achieve our purpose.We propose a human detection dataset labeled by bounding boxes from industrial surveillance videos.This dataset contains 3029 images which includes nearly ten thousand human bodies.We use this dataset to train Faster R-CNN to detect pedestrian.Considering the small size of the object,we modify parameters including the size of the anchors and the threshold in the RPN.We also use Soft-NMS in post-processing because of the intensive of the target.Final results show that we improve the accuracy of the model without retraining the model.We add the indoor dataset into the validation set while it not a part of the training set in order to prevent over-fitting during training process to improve generalization ability.We use human aspect ratios to get workers' head position.Then we use a sliding window to narrow the region of the hardhat in order to classify images with traditional methods or neural networks.We use pedestrian datasets to train and test model.As the experiment shows,the detection method based on Faster R-CNN get great accuracy and generalization ability in factory surveillance videos.Due to the limitations of the marking method and test set,the evaluation of model is not sufficient.Further research should consider to use mask to represent the human body and weight the data in test set to adapt to the detection of hardhat.