Deep Learning Technique for Object Detection from Panoramic Video Frames

Abstract

The objective is to train a YOLOv3 algorithm with necessary enhancements to accurately detect the safety helmet from the video frames which can be used to find the people working in the construction site or riding bike without helmet in the traffic. During the recent past the dominance of deep learning algorithms increased in solving problems in the field of computer vision especially for image classification and object detection. The available algorithms can be divided in to two major categories, 2-stage detection (based on region proposal network) and 1- stage detection. For real time detection of objects from surveillance videos, YOLO based detection is considered to be more suitable approach due to its high speed detection. The loss function and other factors pose few challenges and limitations as the detection accuracy degrades especially when the training dataset is unbalanced. The loss function is modified to overcome the effect of different scale of the object of the same category. This paper utilizes the DarkNet-53 approach, a 53 layered deep convolutional neural network to extract features. The proposed YOLOv3 based safety helmet detector especially the feature extractor is trained on a custom built dataset. The detector achieves a higher detection speed and accuracy with higher generalization ability. The performance of the trained model is tested on panoramic images generated by stitching multiple video frames captured from the surveillance videos. The results demonstrate that the trained model can be utilized to detect the safety helmets from the video frames in real time. The presented approach will be an effective alternate solution for detecting the safety helmets and enhance the safety practices at construction site and road traffic.