Technical Paper
Road Actor Intention Prediction using Video Auto-Encoders
2024-04-09
2024-01-2011
Autonomous vehicles rely heavily on sensors for perception, but mainly in Level 2 autonomous systems, the outputs of the sensors are used in deterministic approach. In this research paper, we propose a more sophisticated method for identifying driver inten tion by utilizing a trainable neural network based on the Transformer architecture and uses a masked Auto Encoder to analyze image sequences within a video. This allows us to predict the intention of target vehicle without the need for explicitly detecting the vehicle or other objects within the sequence using multiple models. The prediction from this model will be input to the general sensor fusion algorithm along with other deterministic parameters, paving the way to reduce the false positives and in some specific cases increase the efficiency of the current active safety functions. This end to end approach is more effective in identifying relevant objects and eliminates any latency that a multi model process could introduce.