In urban roads, pedestrians crossing the road suddenly is very dangerous, with great uncertainty and complexity. At this time, the emergency collision avoidance control of vehicles is particularly important. Therefore an autonomous emergency collision avoidance control that leverages deep Q-learning (DQN) which can learn and adjust accordingly to make realtime braking decisions based on input from integrated image and sensors information was proposed. Considering the varying range of initial vehicle speed and the impact on pedestrian, the reward function is divided into several piecewise functions based on different stopping distances of ego-vehicle. This paper train the policy using the deep Q-learning with this reward function, in order to let the agent to learn how to get as close as possible to pedestrians and stop in place while avoiding collisions with them. Finally the experiments focus on a pedestrian crossing scenario in an urban road and shows that ego-vehicle not only prevents collisions, but also ensures a smooth change in brake application as the vehicle exits the emergency situation, much like human driving.