In order to facilitate the counting of turning traffic flow and to enhance the detection speed and accuracy of turning traffic flow at intersections, a deep learning-based method was suggested for detecting, tracking, and counting turning traffic flow at urban crossings. Initially, the YOLOv5s, which was lightweight and efficient, was chosen as the target detection framework after conducting a comparative analysis. Unmanned aerial vehicle (UAV) aerial photography was utilized to record video footage of traffic movement at urban intersections, resulting in the development of a dataset of vehicle aerial photography photos. The pre-training weights and the most recent weight files were utilized to conduct training and testing on the self-constructed dataset. The model evaluation shows that the vehicle detection model using YOLOv5 exhibits great detection speed and accuracy. The model’s box_loss value declines rapidly and stabilizes at 0.038, while the mAP_0.5 value climbs swiftly and stays near 0.91.After that, the DeepSORT model was used as the backend multi-vehicle tracking technique, and a corner-to-centroid coordinate transformation was used to simplify the extraction of vehicle trajectories. The precision of the driving trajectory line was evaluated thereafter. To improve the robustness of trajectory points’ coordinate information, a corner-point-center-of-mass point coordinate transformation was suggested to tackle the issue of corner points in the detection frame. A sixth-degree polynomial was used to model the vehicle trajectory. Unsuitable trajectory lines were rotated and optimized to meet the function mapping requirements and ensure good fitting of all trajectories. Turning vehicles were detected and counted by using a predetermined threshold to determine the turning angle. Ultimately, to validate the performance of the proposed turning vehicle flow detection method, vehicle detection experiments were conducted at a city intersection as an illustration. The manual counting values were compared and analyzed against the detection results obtained using this method. The results show that the average detection accuracy for the four flow directions is 92.9%, with a maximum of 95.7%, meeting the standard detection requirements for turning vehicle flow in real intersection scenarios.