Moving+: Semantic Scene Classification on YOLOv5

For intelligent moving agents, robots especially, it is important to be aware of the surroundings to help analyze the situation and what might happen in the future. Scene classification is a hotspot with the development of moving agent. Different from the approach directly solving the problem caused by moving platform, our approach does the classification on object detection results from YOLOv5, which is with meaningful and semantic information. Since YOLOv5 works on frames in video, with state of art detect speed and accuracy in area of object detection, it can perfectly avoid performance degradation caused by the moving platform. By further integrating with TF-IDF, five ways to train the model are obtained, the semantic representation sequence is feed into LSTM to handle the temporal relations among frames. Our dataset was consisted with three parts: moving+ dataset: taken from a mobile robot platform, extension dataset: downloaded from internet by keywords retrieval and mixed dataset: mix both of them. Experiment results on three datasets prove the effectiveness of our approach particularly in the moving+ dataset and mixed dataset, our approach shows a high recognize accuracy up to 93% and 92%.