FM-Mnet Real-Time Object Detection Model for Video Feature Fusion Processing

Abstract

Current work on object detection mainly uses the method of one-stage or two-stage. The one-stage method regresses the probability of category and position coordinate value of the object, which results in lower accuracy. The two-stage method firstly generates a series of candidate boxes as samples by the algorithm, secondly, it classifies the samples through the convolutional neural network. Even though it has a high accuracy rate, but the complexity of training time of the model is relatively high. In this paper, we proposed a FM-Mnet model, which is based on the feature fusion method to addresses both limitations. Through this model, a relatively high accuracy rate can be obtained with feature fusion strategy. The experiment shows that this model could lead to a considerable performance in object detection when there are only about five videos in each category.