With the development of intelligent transportation technology and the increasing complexity of traffic scenarios, higher requirements are imposed on real-time vehicle target detection and speed measurement stability in vehicle-road cooperative systems. To address the issues of low efficiency, high computational complexity, and low detection accuracy in complex traffic scenarios encountered by traditional vehicle detection techniques, a lightweight vehicle detection and speed measurement system based on YOLOv5s is constructed. A Multi-scale feature fusion architecture is adopted to achieve the detection of vehicles of different sizes. The Kalman filter algorithm is employed to fuse the temporal information of target center points for estimating vehicle speed. A threefold smoothing strategy, combining median filtering, moving average, and a speed change rate threshold, is introduced to enhance the stability of speed measurement. Experimental results demonstrate that the system achieves an F1 score of 98% for detection accuracy in both self-built datasets and actual video streams, with relatively stable speed measurement. A balance between detection accuracy and real-time performance is realized, providing technical support for vehicle-road cooperation. The system can be applied to small and medium-scale intelligent transportation scenarios.