In autonomous driving, we know that the behaviour prediction module is concerned with understanding how the agents in the environment will behave. Similarly, in the perception module, the tracking algorithms are responsible for getting an estimate of the object's state over time.
Although both processes might be doing estimations (because data sources aren't perfect and/or have noise), there is a key difference:
Object tracking cares solely abouth where objects are now. That means that there is actual sensor data that can support the current position. For example: from a camera and a lidar, the computer predicts where a vehicle stands.
Trajectory prediction is done with the main purpose of predicting where objects will be in the future, meaning that there is no sensor data yet. For example: from past data, the computer predicts where a vehicle will be after 1 second from now.
However, both processes might need each other. In order to predict a trajectory, it could be easier to work with curated positions given by an object tracker than raw sensor data. On the other hand, past predictions might be given as input to an object tracker, in order to mitigate noise in the sensor data and acting so as a belief-based filter.