This paper introduces an image-priority approach with integrated 2D-3D views for continuous multi-target tracking and speed estimation in camera-LiDAR surveillance systems. It addresses differences in acquisition frequency， resolution， and viewing angles between camera and LiDAR-based speed estimation tasks. Geometrically matched feature points are selected， and external parameters are computed using the Direct Linear Transformation method for online calibration between devices. A vision-guided， frustum-based spatial method combines 2D and 3D localization， uses high-resolution ground points to define area boundaries， and adapts clustering parameters in a 2D top-down view transformed from a 3D frustum perspective. This approach helps eliminate irrelevant points and address mixed-resolution point cloud detection due to varying viewing angles. The speed estimation process employs Kalman filtering and vehicle motion states， modeling speed estimation as observation equations using discrete synchronized frame point cloud data. The observation noise covariance matrix is calculated based on point cloud resolution， allowing continuous optimal estimation and reducing observation noise and asynchronous timing effects. Experiments on traffic scene datasets show that the method achieves an average absolute error of 0.276 4 m/s and a root mean square error of 0.325 1 m/s， with a maximum detection range of 103.211 m， demonstrating high accuracy and practicality.