An intersection traffic control method, apparatus and system are provided. The method includes that: a vehicle signal of a first vehicle at an intersection and a vehicle signal of a second vehicle located in a set zone in proximity to the intersection are acquired; the vehicle signal of the first vehicle and the vehicle signal of the second vehicle are input into an instruction learning model trained in advance based on a reinforcement learning principle, and a score of a preset traffic indicator of the first vehicle after executing a respective candidate action instruction is calculated; a reward of the first vehicle when executing the respective candidate action instruction is acquired according to the score of the preset traffic indicator, a candidate action instruction corresponding to a maximum reward is determined as an output result of the instruction learning model, and a next action instruction is determined according to the output result; and navigation of the first vehicle through the intersection is controlled according to the next action instruction.