Traffic lights operating on a fixed schedule are mostly time-consuming; for example, running green signals in the absence of vehicles, leading to a buildup of long queues at red lights. This inefficiency results in congestion in cities, contributes to delays and economic losses and intensifies pollution levels. In this study, a deep learning-based adaptive image processing traffic light control system for real-time dynamic regulation of signals was proposed. Different from typical sensor-based solutions, the proposed method uses established surveillance cameras, enabling cost-efficient deployment and easy installation. A YOLOv10-based detection model identifies and classifies vehicles by type, applying weight factors to effectively estimate traffic demand. A dynamic timing algorithm enables continuous redistribution of green-light durations due to existing unbalances in the flow for any or all intersection phases. A practical microcontroller-based system might be integrated directly into the existing infrastructure. For assessment, the model used data from 12,500 images labelled accordingly and divided into the following: 70% for training, 15% for validation and 15% for testing. The model was assessed in a SUMO-based simulation of a very busy four-way intersection and actual deployment in Baghdad, Iraq. Compared with fixed time control, this adaptive system reduced vehicle wait time by up to 91.7%. Furthermore, results indicate reduced fuel consumption and CO2 emissions, thereby leading to considerable economic and environmental benefits. Overall, the proposed framework represents a practical and scalable implementation for modern traffic management, overlooking possible implementations of enhancements such as prioritisation of emergency vehicles and multi-intersection coordination.