Intelligent traffic gesture recognition based on YOLOv8 algorithm optimization and OpenPose

To address the challenges of target loss due to extreme lighting interference, dynamic occlusion, and misjudgment of traffic gestures in complex traffic environments, this study proposes an intelligent traffic gesture recognition method that integrates an improved "You Only Look Once version 8" target detection algorithm with lightweight OpenPose pose estimation. First, an adaptive detection framework integrating physical optics preprocessing and optical flow motion compensation is constructed to achieve zero-latency target tracking. Second, a lightweight OpenPose based on a deep separable convolutional backbone network and a serial feature fusion mechanism is proposed for low-level skeleton extraction. Furthermore, a long short-term memory network and a dynamic time warping algorithm are dynamically introduced into the action classification decision layer for temporal feature interpolation and nonlinear optimal path alignment. Results show that in highly reflective scenes, the improved target detection algorithm achieves an average accuracy of 87.60% and a false detection rate of only 7.05%. The lightweight pose estimation algorithm achieves a correct keypoint ratio of 93.8% and an edge-end inference frame rate of up to 58 FPS. This demonstrates that the intelligent traffic gesture recognition method effectively overcomes physical visual interference and edge computing power bottlenecks in complex road environments, providing reliable technical support for intelligent vehicle decision-making in complex traffic scenarios.