Research and Application of Small Object Recognition Based on TX2 in the Field Scene

摘要

In the field environment target detection, there are problems such as low grab rate due to fast moving objects and poor quality of acquired image data. Considering these issues, this paper designs a field small target detection system that is mounted on the TX2 embedded development board. To solve the problem of poor image data quality, a super-resolution method DRCN (deep-recursive convolutional network) is introduced to enhance the data clarity. Besides, the structure of DRCN is set to be deeper. To handle the problem of low capture rate of fast moving targets and the problem of memory shortage in TX2, a new method based on increasing inter-frame difference threshold judgment is employed. The new method namely the three-frame difference method based on difference threshold judgment can filter the "blank" differential image. Differential mapping preservation only when fast moving targets occur.It can improve the capture rate of fast moving targets and save storage space at the same time. Experiments show that the designed system can adapt to TX2 and effectively improve the effect of field environment detection.