Towards Efficient Video Object Detection on Embedded Devices

The challenge of adapting various object recognition techniques from still images to videos remains unsolved. When applied to videos, methods that are specifically designed for images do not perform well due to several complications. These include blurriness, shifting or ambiguous locations, subpar quality, and other similar concerns. In addition, a lack of effective long-term memory in video object detection has yet to be addressed. It is widely recognized that consecutive frames in a video tend to produce highly similar results in most cases. Therefore, this characteristic can be exploited to improve performance. Moreover, the information contained in a series of sequential or non-consecutive frames exceeds that of a single frame. In our research, we have introduced a novel recurrent cell for feature propagation and have determined the optimal layer placement to augment the memory span. This has resulted in superior precision compared to methods presented in previous research. Furthermore, hardware constraints may exacerbate this issue. Therefore, we have focused on implementing and improving the effectiveness of these techniques on embedded devices. Our approach has yielded impressive results, with a 67.5% mAP accuracy on the real-time ImageNet VID dataset for mobile devices at a rate of 62 fps.