Research On Target Height Measurement Technology Based On Deep Learning

In order to improve the safety and comfort of car driving,many car manufacturers focus on the design and development of Advanced Driver Assistance Systems.With the development of artificial intelligence,autonomous driving technology has become the latest development direction of the current automotive industry.Environmental perception is the foundation of autonomous driving technology,in order to improve the vehicle's obstacle avoidance capabilities.Advanced Driver Assistance Systems needs to have the capability of target detection,ranging and target scale measurement.At present,self-driving cars mainly rely on sensor devices such as lidar to detect obstacles around the car,but in good traffic conditions,self-driving cars can also use the camera to implement vision-based object detection and ranging tasks,this method has lower measurement costs and can be used as an auxiliary method for cars to achieve environmental awareness.To further improve obstacle avoidance of autonomous vehicles,this paper studies object detection,ranging and target scale measurement respectively.Based on these three technologies,a binocular stereo vision system for height measurement of "person" is designed in this paper.The stereo matching algorithm is one of the important techniques to achieve distance measurement using a binocular camera.This paper designs a stereo matching algorithm based on PSMNet.First,this paper uses the channel attention mechanism to improve the cost volume in PSMNet.Specifically,this paper uses three different distance metric functions to calculate channel weights,after experimental comparison,this paper chooses to use Pearson correlation coefficient to improve the matching cost volume.Secondly,this paper designed the Spital structure and the Encode-decode structure to improve the 3D CNN in PSMNet.After experimental comparison,this paper chose the Encode-decode structure as the improvement solution.This paper designs a target detection algorithm for "person" based on Yolov3.In order to improve the accuracy of the detection box,this paper redesigns the anchor,and uses Io U confidence and soft-NMS to improve the post-processing.For purpose of achieve multi-scale object detection,this paper introduces dilated convolution in the FPN structure of Yolov3.In order to further improve the detection speed,this paper uses Mobilenet?v2 to optimize the backbone of the network.Through a series of improvements,the optimized object detection algorithm has higher detection accuracy and faster detection speed than Yolov3.To further improve the accuracy of the detection box,this paper corrects the bottom of the detection box based on the edge map and corrects the top of the detection box on the disparity map.It can be found that the corrected box has higher accuracy and smaller fluctuations.Based on the stereo matching algorithm,object detection algorithm and detection box correction algorithm,and combined with the conversion relationship between the pixel coordinate system and the camera coordinate system in stereo vision,this paper implements a vision system for measuring height of person.