Real-time Indoor Object Detection Based on Deep Learning and Gradient Harmonizing Mechanism

Due to the indoor environment is complicated, and the proportion of the background is much larger than the object, the positive and negative categories of the sample are not balanced. This paper proposes that the imbalance of sample categories is due to the imbalance of difficult and easy samples which can be reflected by the gradient norm of the sample. The gradient density is introduced to solve the uneven distribution of gradient norm. The parameter "gradient density"is added to the Yolov3 loss function which is one of the best one-stage object detection frameworks that can balance accuracy and speed in order to improve the sample category imbalance. Considering the high computational complexity of the sample gradient density, the gradient norm is divided into equal-width intervals, and the same gradient density is adopted for the samples falling in same region so as to simplify the calculation and improve the training efficiency. The experimental results demonstrate that the improved approach achieves best detection accuracy and obtains the most accurate bounding boxes of indoor objects to be classified with quick speed. Therefore, the gradient harmonizing mechanism of samples can improve the sample category imbalance. The improved indoor object detection algorithm can be applied in the field of intelligent indoor monitoring, indoor service robots.