Smart Vehicle Accident Detection System using Faster R-CNN

A new dataset is being provided by the paper for analyzing traffic incidents. The mission is about addressing lack of publicly available data needed for research into automated spatiotemporal annotations for road safety. Because of the object sizes and sophistication of scenes, we found a substantial degradation of detection of object inside the pedestrian group in our dataset after analyzing it. We are using here 2 datasets one of them is the DETRAC dataset which is used for vehicle detection and the other one is CADP which contains the accidents and we are supposed to detect them. Car Accident Detection and Prediction (CADP) dataset contains YouTube video segments 1,416 around, of which 205 have absolute spatiotemporal annotations. Due to the object sizes and sophistication of the scenes, we found a major degradation of object detection in the pedestrian group of the CADP dataset. To this end, we suggest incorporating Augmented Context Mining (ACM) into the Faster version of R-CNN detector for the improvement of small pedestrian detection accuracy.