An in-vehicle violence detection system includes a speech and non-speech audio recognition module capturing occupant threat words and non-speech audio events. In-vehicle accelerometers generate data analyzed in a shaking movement recognition module. A heart rate and breathing rate detection module measures physiological changes in occupant heart rates and breathing rhythms. An in-vehicle semantic scene recognition module analyzes occupant non-verbal interactions. Occupant threat indicators including an audible threat indicator are generated by the speech and non-speech audio event recognition module. Visual threat indicators are generated by the in-vehicle semantic scene recognition module. Physiological threat indicators are generated by the heart rate and breathing rate detection module. Vibration-based threat indicators are generated by the shaking movement recognition module. The four threat indicators are fused to estimate a threat level, consolidated by incorporating contextual information including a vehicle location and time. Different actions are taken based on the consolidated threat level.