To meet the traffic control requirements under different traffic accident scenarios, a method for classifying traffic accident severity and implementing control strategies based on residual road capacity was proposed. The concept of residual road capacity was introduced to enable real-time accident severity assessment. To address the reduced target recognition accuracy caused by haze and vehicle occlusion in accident scenarios, a coordinate attention mechanism was integrated to enhance YOLOv5, enabling precise detection of traffic flow at the accident point and upstream sections. To overcome the limitation of insufficient real-world samples for YOLOv5, VISSIM microsimulation was employed to supplement low-sample scenario data. Combining Video detection and simulation data, polynomial fitting with the akaike information criterion was used to construct the residual capacity curve, from which the residual capacity coefficient was derived. Based on this coefficient, traffic accident severity levels were classified, and differentiated control strategies were developed. The effectiveness of the method was validated through VISSIM simulation. Results show that the method enables accurate and dynamic accident severity classification. The implemented control strategies effectively reduce the impact of accidents on road capacity. Under first-level and third-level incident responses, congestion length is reduced by over 95% compared to uncontrolled scenarios, significantly lowering the risk of secondary accidents and network congestion. This method provides reliable technical support for emergency traffic control on highways.