Traffic congestion adversely affects the quality of life of urban residents, wasting hours and causing daily inconvenience. Alongside this, it causes severe environmental issues. Clearly, improving traffic signal plans can positively mitigate this challenge and reduce its adverse impacts. In the paper, a novel, practical deep reinforcement learning agent is proposed for intelligent traffic signal control. In contrast to existing literature, which assumes the availability of individual or aggregated intersection vehicle metrics such as position, waiting time, and speed, our proposed agent state utilizes only a finite image sequence. This sequence can be collected by a single camera that captures the intersection bird’s eye view once per second. The image sequence is processed by a convolutional neural network, which extracts both spatial and temporal traffic patterns without any handcrafted feature engineering. The proposed state and model are integrated in a Deep Q-Network (DQN) agent that interactively manages the intersection signal. Experimental work in a real-world scenario from the RESCO benchmark revealed that the proposed agent surpassed two baselines: a fixed-time policy and another feature-based DQN agent. Compared to the baselines, the presented agent improves several metrics, including time loss, trip time, waiting time, and queue lengths.