Vehicle-mounted monocular visual velocity measurement method based on MaxViT and improved geometric feature point method
Vehicle-mounted visual speed measurement technology,as an important component of composite speed measurement technology for autonomous vehicles,is characterized by low hardware costs,strong algorithm scalability,and accurate measurements at low speeds,offering broad application prospects. To further improve the accuracy and robustness of visual speed measurement algorithms under various working conditions,this article combined the advantages of geometric feature point methods for high accuracy when sufficient feature points were available and deep learning methods for stable speed measurement in multiple scenarios. A vehicle-mounted monocular visual speed measurement algorithm was propose based on MaxViT and an improved geometric feature point method. This algorithm constructed a dual-channel,parallel processing system based on a dual-input MaxViT network and an improved geometric feature point method. This could continuously process a sequence of three input images obtained from the forward-facing camera of the vehicle. It estimated the current vehicle speed in a rolling manner. The dual-input MaxViT network differentially extracted optical flow features from different regions of the images,and estimated the velocity interval with a confidence level of 90％ where the current velocity lied. The improved feature point method could calculate the current velocity estimate based on the motion of feature points. If the velocity estimate falls within the velocity interval estimated by the dual-input MaxViT network,it was used as the real-time vehicle velocity measurement value. Otherwise,the midpoint of the velocity interval was used as the real-time vehicle velocity measurement value. After running the algorithm for multiple frames,the midpoint of the velocity interval was used as the velocity output for the current frame to reduce cumulative errors. Experimental verification was conducted using a self-built dataset containing six different vehicle velocity,including velocity below 40 kilometers per hour and various driving scenarios such as acceleration,deceleration,and straight and curved roads. The theoretical velocity accuracy of GPS velocity signals was used as a reference velocity with an accuracy of 0.1 m/s. The proposed method can achieve an average relative velocity measurement error of less than 1.37％ and a maximum relative velocity measurement error of less than 6.13％. The experimental results demonstrate that the proposed approach effectively enhances the accuracy and robustness of vehicle-mounted visual velocity measurement,providing theoretical support for the integration of diverse vehicle-mounted visual velocity measurement methods.
vehicle velocity measurementvisual velocity measurementdual-input-MaxViT networkcharacteristic point methodverification output
万方数据
维普
