10 June 2024 in IEEE Sensors Journal
In the intricate urban road scenarios, the algorithm for vehicle speed measurement based on image sensors encounters challenges of low precision. Often, a confluence of video and radar solutions is employed, impeding large-scale applications in the engineering domain. By amalgamating deep learning algorithms and camera imaging principle, a speed measurement algorithm was crafted that integrates scene and object features. Initially, the YOLOv5s object detection model was employed, and refined the model’s backbone and loss function to elevate the precision of vehicle detection. Subsequently, in conjunction with the RAFT algorithm and camera imaging principle, the geometric relationship between road scenes and images was established, deriving vehicle speed based on the coordinates of vehicle’s feature points. Furthermore, a vehicle dataset was curated to train and assess the improved YOLOv5s model. Finally, employing an experimental vehicle traversed the designated video detection section at five constant speeds, to validate the efficacy of the speed measurement algorithm. Experimental results illustrate that, in comparison to the original model on the self-established dataset, the improved YOLOv5s demonstrates enhancements of 1.7% and 1.4% in mAP@0.5 and mAP@0.5:0.95, respectively. The mean absolute error and root mean square error for all speed measurement values are 1.95 km/h and 2.38 km/h, respectively. In summary, this algorithm is applicable to vehicle speed measurement tasks in complex traffic environments, showcasing better precision.