FCOS Based Seatbelt Detection System Using Thermal Imaging for Monitoring Traffic Rule Violations

This work is related to building a Seatbelt Detection system based on Fully Connected One Shot (FCOS). It is one of the most recent Deep Learning approaches primitively built using single shot detection proposal. Unlike the double stage region-based object detection schemes this technique do not follow semantic segmentation, it does not undergo loss of the object information such as disappearance of the gradients and it does not require pre-defined anchors. This technique comprises strong feature extractors and reinforce multi scale object detection and it is very quick in the multithreaded GPU environments. Since our fundamental research is concentrated on object classification related to ADAS applications, as a first step we choose to detect the drivers and co-passengers of a four-wheeler vehicle not wearing seatbelts from thermal dataset. Therefore, we used thermal images and videos possessed from thermal cameras 5m to 20m away from the vehicles as our dataset in building the model and testing. The FCOS extracts the features of an object using its efficient per-pixel fashion. Finally, the performance analysis of these model in terms of mean Average Precision (mAP) indicates that the modelling using FCOS performs in a promising way and it can be used in automatic seatbelt detection systems.