Study of Detection Method on Real-time and High Precision Driver Seatbelt
Dongsheng Yang, Ying Qin Zang, Qingshan Liu · 2020
The CPU-based deep learning seatbelt detection cannot satisfy the real-time requirements. The precision of seatbelt detection obviously decrease on the basis of deep learning due to the influence of some factors, like the light, rain and snow, window reflection etc. In order to improving detection precision under running CPU in real-time, the SSD MobileNet V2 network is simplified from network pruning and quantitative processing in this paper. The precision and speed of the detection have been improved according to the above optimization. The precision of seatbelt detection can be improved utilizing the regression constraint of seatbelt within the driver after integrating into one classifier. After that, the feature augmentation was added into the samples in order to improve the precision of seatbelt detection in deeply. Finally, the same driver was tracked by the particle filter algorithm. If the tracking failure, the obtained driver feature vector of deep learning will be used to instead of the histogram feature of particle filter and used for the following multi-frame image detection of seat belt. Finally, the detection results are 95.21% for detected precision, 97.19% for recall rate, and 21FPS for detected speed, respectively.