Object Recognition And Positioning For Automatic Driving Collision Avoidance Warning

In recent years,with the rise of artificial intelligence,autonomous driving technology has attracted more and more people's attention,among which computer vision is one of the key technologies.In the automatic driving scene,the car has extremely high requirements for the detection of obstacles around it.In order to enable the automatic driving system to recognize the object type,determine the shape of the object,and measure the distance of the object,this paper proposes a method based on Mask R-CNN and binocular vision.The object recognition and positioning method uses the binocular image to accurately locate the object,and then uses any image in the binocular image to identify the object type.These two methods can be calculated in parallel under the GPU and reach a high level.Accuracy and quasi-real-time requirements.At present,object recognition s ystems usually use feature pyramids(such as FPN)to learn multi-scale representations to obtain better results.However,current feature pyramid designs still cannot effectively integrate semantic information of different scales.This paper proposes a new configuration framework(NFPN)for the current feature pyramid,combining low-level representation with high-level semantic features in a highly nonlinear and efficient manner,using global scanning to emphasize the global information of the complete imag e,and then performing the region.Reconfigure to model local patches within the region of interest that can span different spatial locations and scales to better identify different sized object categories.The experimental results in and are 38.2 and 39.1,respectively.A leading stereo matching algorithm(such as the VariableCross method)uses a fixed color threshold and a rigid cross to construct a support window.However,this construction method has difficulty in processing images of different topography.In response to this problem,this paper proposes a new stereo matching algorithm(DAS method)that uses a dual adaptive support window.The method uses a support window adaptive to the appearance and shape of the image area to perform cost aggregation,and combines the color absolute difference with the cost initialization of the census transformation,multi-directional scan line optimization and parallax refinement to achieve an accurate stereo matching.system.The results show that the latter's average parallax error is 28.15% lower than the former.