Research On Lightweight Deep Learning Vehicle Detection Algorithm And Collision Warning Technology

As an indispensable means of transportation nowadays,vehicles bring convenience to people,but also bring a lot of safety risks.Unconsciously following the vehicle too close in the process of driving often leads to the driver’s failure to respond in time when an emergency breaks out in front of it,thus causing traffic accidents,bringing economic losses to people and possibly causing traffic congestion in road networks,or even causing a number of casualties.As the key technology of Advanced Driving Assistance System(ADAS),Forward Collision Warning(FCW)enables the driver to perceive the danger in advance through targeted warning at a reasonable time,so as to reduce or avoid the occurrence of collision accidents.Therefore,FCW is of great significance to road traffic safety.In view of the high false alarm rate,poor real-time performance,high computational complexity,and high cost of the existing FCW system,this thesis uses lightweight vehicle detection algorithms to reduce the complexity of the entire system to ensure the real-time performance of the system.By fusing various information such as the distance,vehicle speed and collision time of the preceding vehicle to improve the accuracy of the early warning,high-precision,low-latency,multi-level vehicle collision early warning is finally realized.The thesis starts from the following two aspects:First of all,due to the limited computing resources and storage space of in-vehicle terminals in real scenes,most deep learning methods have problems such as high computational complexity,poor real-time performance,and large model occupation of memory,which make it impossible to apply them to actual driving scenarios.Based on this,this thesis proposes a lightweight vehicle detection algorithm based on deep learning.The algorithm uses the idea of designing a lightweight network model to design a lightweight backbone feature extraction network-My-Mobile Net,and build it into YOLOv4 target detection.In the model,the vehicle detection complexity is reduced,the size of the model is compressed,and the real-time detection is improved while ensuring the accuracy of the detection.Through training on a variety of different types and different scale data sets,the robustness of the detection algorithm is improved.Finally,through the simulation analysis of the data set collected in Xi’an and the data set shared on the network in the scene with low computing power,it is finally verified that the model size is greatly compressed under the condition of meeting the real-time demand,and it has a high accuracy rate in a variety of scales and scenarios.Secondly,in actual driving scenarios,due to the singleness of the traditional TTC(Time To Collision)model measurement indicators,when the front vehicle and its own vehicle are not in the same lane,a large number of false alarms will occur,which will affect the normality of the driver.drive.Therefore,based on actual needs,this thesis proposes a vehicle collision warning algorithm based on multi-data fusion.First,the Deep-SORT algorithm is used to track the vehicle in front,and then the horizontal and longitudinal distance between the vehicle and the vehicle in front are calculated in real time according to the geometric model,and finally the relative speed,collision time,horizontal and longitudinal distance between the vehicle and the vehicle in front are used.Data fusion and reasonable setting of early warning thresholds will eventually divide the collision risk into two levels.It is compared with the traditional TTC model to verify the effectiveness of the above algorithm.