With the advancement in technology, autonomous vehicles will have a great impact on the near future and share the road space with human beings. For ensuring safer navigation, a robust collision mitigation system becomes mandatory which in general is deployed with multi-modal approaches. In this paper, a novel heuristic unimodal approach based on the vision system is presented to estimate the collision priority of vehicles on road. The priorities are estimated from the perspective of an ego vehicle that may be equipped with a vision based driver-assist system or a fully autonomous vehicle. Crowd-sourced videos from YouTube involving vehicular collisions captured by the dashboard camera of vehicles are considered. To detect the moving vehicles in the video, a deep learning-based pre-trained object detection model and a tracking algorithm are used. From the bounding box output, an estimate of collision priority of detected vehicles concerning the ego vehicle is obtained using an empirical heuristic-based approach. A simpler collision warning and navigation suggestion is also incorporated as a credible advanced driver assistance system element. The proposed qualitative approach performs well for input videos and a more robust estimate can be achieved by combining with other quantitative semantics of traffic parameters.
Estimation of Collision Priority on Traffic Videos using Deep Learning
2020-12-20
428860 byte
Conference paper
Electronic Resource
English
Traffic Analysis on Videos Using Deep Learning Techniques
Springer Verlag | 2023
|Traffic signal priority control based on shared experience multi‐agent deep reinforcement learning
Wiley | 2023
|Unsupervised learning from videos using temporal coherency deep networks
British Library Online Contents | 2019
|Traffic signal priority control based on shared experience multi‐agent deep reinforcement learning
DOAJ | 2023
|