Motion Autonomy and Safety issues in Autonomous Vehicles are strongly dependent upon the capabilities and performances of Embedded Perception and Situation Awareness systems. Recent benign and severe accidents (e.g. Tesla or Uber) have shown that the level of safety obtained using currently tested autonomous driving systems is still insufficient. This speech addresses this important perception and safety issue, and presents how it can be addressed using Bayesian and Machine Learning approaches. The approach is illustrated using results obtained in the scope of several Research and Development projects conducted in cooperation with the French IRT Nanoelec and with several industrial companies such as Toyota and Renault.