Research On Vehicle Detection And Tracking Algorithm Based On Machine Vision Fusion

The intelligent vehicle environment awareness system obtains the information about the road surrounding the vehicle through various type of automotive sensors to provide a reference basis for subsequent vehicle decisions.Among the mainstream methods of fusion of millimeter wave radar and vision sensors,the deep learning based vision detecting algorithm model is relatively complex and the parameter quantities are large,compromising the real-time capability of target detection;millimeter wave radar can’t provide height information and is vulnerable to false detections and omissions in complex and congested scenarios.This paper proposes a decision level algorithm for information integration,respectively for vehicle detection and tracking using millimeter wave radar and single vision camera,and optimizes the current studies in three areas:detection and tracking by visual sensors,filtering and tracking by radar targets,and sensor information fusion strategies,with the following main research elements.(1)For the YOLOX convolutional neural network with high detection accuracy but complicated model with slow detection speed,a CBS module is substituted for a Resblock Body module in the backbone network to reduce the complexity of the model;the SE attention mechanism is added to the Shuffle Net V2 base module and this structure is substituted for the CSPLayer module of the neck network;the CBAM attention module is added to the three feature layers of the backbone network output to improve the model detection performance and the new convolutional neural network is being trained with the KITTI dataset.The deep residual network replaces the Deep Sort original re-identification network,and the optimised Deep Sort multi-target tracking algorithm is presented.The test results show that the proposed detection and tracking the algorithm to improve the stability of visual detection,with real-time and accurate detection effect.(2)According to the millimeter wave radar selection,the raw radar data is first parsed and data processed to filter out stationary targets and spurious targets.A kinematic model of a uniformly accelerated vehicle is developed and a multi-objective tracking algorithm is proposed based on square root Cubature Kalman filtering and joint probabilistic data correlation.The simulation shows that the tracking algorithm proposed achieves continuous and accurate tracking.(3)Establishing decision-level information fusion algorithms for millimeter wave radar and vision sensors.The sensor spatial fusion is accomplished through the coordinate conversion relationship,and the sensor temporal fusion is accomplished by using the millimeter wave radar sampling period as the sampling period after fusion.The vision sensor and millimeter wave radar fusion strategy is established,and the corresponding radar and vision sensor fusion results are output according to the different overlapping areas of the area of interest generated by the millimeter wave radar projection and the detection area of the vision sensor.Finally,a real vehicle validation platform is built and real road data is collected to process the road data offline.(4)As a conclusion,in a well-illuminated,structured interurban road environment,the visual detection and tracking algorithm proposed in this paper effectively improves target detection accuracy and the real-time performance,and the decision-level information fusion algorithm can achieve accurate and stable detection and tracking when multiple vehicle targets are in front of this vehicle,significantly reduces the vehicle detection and tracking leakage and false detection rate,and has good environmental adaptability in a variety of road scenarios.