In low light conditions, images tend to become blurry and dim, and vehicles and other traffic signs become less recognizable. This makes vehicle detection algorithms more difficult, increasing the likelihood of false and missed detections. For this reason, a real-time automatic vehicle flow detection method is proposed for road intersections under low light conditions. Multi-dimensional attention mechanism and recursive pyramid network are embedded in U network architecture, and combined with regional feature aggregation network, the regions of interest in low-light intersection images are screened. Based on this, the deep reinforcement learning model is used to make the detection frame fit the target vehicle, and the complete detection frame fit the vehicle is obtained through the regression fine adjustment of the multi-layer fully connected network, and the real-time automatic detection of vehicle flow is completed. The experimental results show that the proposed method has accurate and effective detection ability for vehicles, reflective objects and luminous objects in the distance and dark under various weak lighting conditions, and provides a reliable reference for traffic control.