3D Vehicle Recognition And Distance Measurement Based On Monocular Vision

In recent years,car owners have increased the requirements for the practicality and richness of the intelligent driving function of the car.It is very crucial to use sensors to perceive the complex environment around the vehicle.The vehicle 3D recognition and ranging technology in multiple scenes can comprehensively obtain the parameters of distance,pose and shape of the target in space.On the one hand,it provides available data for downstream functions,on the other hand,it improves the practicability of functions and the richness of driving experience.The recognition technology in complex scenes has increasingly become one of the research hotspots.However,it is expensive to have a vehicle identification system in a car for two reasons.On the one hand,most of the solutions of perception technology are implemented by sensor fusion methods,and radar is expensive;On the other hand,the perception algorithm has a high requirement on the computing power of the system chip.However,our chip technology has been monopolized by foreign enterprises for a long time,and the cost of using highperformance chips is too high.In view of this,aiming at the above pain points,the thesis will adopt the low-computing car-scale Soc chip T5 as the carrying platform,and the lowcost wide-angle camera as the only sensor to develop a vehicle recognition algorithm that includes two functions of three-dimensional target recognition and single visual distance measurement.In addition,the network and model used will be generalized and lightweight processing,so that it can run on the T5.Thus forming a set of high cost performance,high accuracy of vehicle identification system.In this system,the 3D target recognition function can restore the target in the forward visual field of the vehicle to the stereo image,and accurately display the position of the target on the image,the front direction and the size of the vehicle.The single visual range function estimates the theoretical distance between the target and the vehicle based on the target orientation estimation and the camera imaging principle.In addition,Nano Det,an object detection network of anchor-base class,was used in the thesis to train vehicle object detection models with AP50 up to 81.8% based on samples and tags of public data sets.In addition,in the thesis,the convolutional neural network Mobilenetv2 is transformed with multiple tags by adding classifier branches,and the target threedimensional parameter regression model is trained based on the samples and tags of the public data set.Finally,the complex scene recognition algorithm based on monocular vision is transplanted to the T5 Soc chip.In the thesis,based on the MNN reasoning framework and OPENCL acceleration,Nano Det and Mobilenetv2 networks run simultaneously in the system with low computing power.The test shows that when all functions run on the embedded device,the processing speed of single frame image is 5FPS.The threedimensional detection function of the vehicle target fits the real target well in different scenes.The average relative error of single visual distance measurement function is7.48%,and the average relative error of each distance section is controlled within 10%,which meets the actual engineering requirements.On the low cost and low computing power domestic chip,the vehicle 3D recognition and ranging function is realized based on monocular vision imaging.