随着智能技术的发展，车辆智能化是当今汽车发展的主要方向。车辆前向碰撞预警（Forward Collision Warning，FCW）是车辆高级辅助驾驶系统（Advanced Driving Assistance System，ADAS）的关键技术，FCW系统通过车载感知传感器来检测自身车辆行驶路径前方的车辆，能在感知到危险时快速向驾驶者传达信号从而减少或避免碰撞事故的发生，符合当今车辆智能化的发展趋势。传统的前向碰撞预警系统一般采取激光雷达进行信息获取，然而这种方法有着系统复杂、成本高和精度低等缺点。近年来，随着计算机视觉图像技术的不断进步，通过摄像头获取道路信息已经变得更加可行。本研究以行车记录仪视频为数据基础，设计一种基于改进YOLOv5（You Only Look Once）算法的车辆前向碰撞预警系统，该系统运用轻量级主干网络来减少计算的复杂度，以确保系统具有实时性。它将自身车辆和目标车辆的车距、安全距离和碰撞时间（Time To Collision，TTC）等多种信息融合起来，提高预警的准确度。最终实现一个多层级、高精度、低延迟的车辆碰撞预警系统。本文主要研究内容如下：

（1）构建基于行车记录视频的车辆图像样本库。视频图像数据主要来源于自行采集数据和部分BDD 100k（A Large-scale Diverse Driving Video Database）数据集中数据，依据车辆特征检测难易程度将车辆分为4类：Car类、Bus类、Truck类和Bike类。设计半自动标注方法，准确标注信息，构建车辆检测数据集和车辆预警测试数据集，并进行数据集特点分析和数据集划分，为车辆检测模型和碰撞预警模型提供数据支持。

（2）构建基于YOLOv5网络的车辆检测优化模型。该模型使用PP-LCNet为主干特征提取网络，并使用Shuffle_Block模块代替YOLOv5中的C3模块，同时剪除一些冗余操作，去掉了Focus模块。同时根据自制数据集重新聚类检测锚框，优化YOLOv5的位置损失函数，形成轻量级的YOLOv5网络模型。

（3）设计基于改进的YOLOv5车辆碰撞预警策略。基于含相机俯仰角的传统几何测距法，设计含相机俯仰角和横滚角的改进几何测距法，完成单目摄像机测距功能。在前方车辆检测与测距的基础上，利用本车与前方车辆的车距，安全距离、TTC等数据对预警阈值进行设定，设计基于最小制动安全距离和相对碰撞时间融合的车辆碰撞预警策略。

（4）开展车辆碰撞预警系统移植与实车试验。以Nvidia Jetson TX2开发板为载体，将改进后的YOLOv5车辆检测算法进行移植。根据本文构建的车辆碰撞预警策略，设定了静止、低速、动态等三种不同试验场景。试验结果表明本文设计的基于改进YOLOv5目标检测算法的车辆碰撞预警系统在各种试验环境下都具有较好的预警效果。

With the development of intelligent technology, vehicle intelligence is the main direction of today's automotive development, Forward Collision Warning (FCW) is a key technology of Advanced Driving Assistance System (ADAS), FCW system through the vehicle The FCW system detects vehicles in front of its own vehicle path through on-board sensing sensors, and can quickly communicate signals to the driver when danger is detected to reduce or avoid collisions, in line with the development trend of today's vehicle intelligence. Traditional forward collision warning system generally takes radar for information acquisition, however, this method has the disadvantages of complex system, high cost and low accuracy. In recent years, with the continuous progress of computer vision image technology, it has become more feasible to obtain road information by camera. In this paper, we design a vehicle forward collision warning system based on the improved YOLOv5 algorithm using the video of a car recorder as the data base, which uses a lightweight backbone network to reduce the complexity of computation to ensure that the system is real-time. It fuses multiple information such as distance, safety distance and Time To Collision (TTC) of own vehicle and target vehicle to improve the accuracy of early warning. Finally, a multi-level, high-accuracy and low-latency vehicle collision warning system is realized. The main research contents of this paper are as follows:

(1) Constructing a vehicle image sample library based on driving record video. The video image data mainly comes from the self-collection data and part of the BDD 100k dataset data, and the vehicles are divided into four categories according to the difficulty of vehicle feature detection: Car category, Bus category, Truck category and Bike category. The semi-automatic labeling method is designed to accurately label the information, construct vehicle detection dataset and vehicle warning test dataset, and perform dataset feature analysis and dataset division to provide data support for vehicle detection model and collision warning model.

(2) Construct a vehicle detection optimization model based on YOLOv5 network. The model uses PP-LCNet as the backbone feature extraction network, and uses Shuffle_Block module instead of C3 module in YOLOv5, while cutting out some redundant operations and removing the Focus module. Meanwhile, the anchor frame is re-clustered and detected according to the homemade dataset, and the position loss function of YOLOv5 is optimized to form a lightweight YOLOv5 network model.

(3) Design a vehicle collision warning strategy based on improved YOLOv5. Based on the traditional geometric ranging method containing camera pitch angle, the improved geometric ranging method containing camera pitch angle and cross roll angle is designed to complete the monocular camera ranging function. On the basis of the front vehicle detection and ranging, use the vehicle distance between this vehicle and the front vehicle, safety distance, TTC and other data to set the warning threshold, and design the vehicle collision warning strategy based on the minimum braking safety distance and relative collision time fusion.

(4) Carry out vehicle collision warning system transplantation and real-world testing. The improved YOLOv5 vehicle detection algorithm is ported with the Nvidia Jetson TX2 development board as the carrier. According to the vehicle collision warning strategy constructed in this paper, three different test scenarios such as stationary, low speed and dynamic are set. The test results show that the vehicle collision warning system based on the improved YOLOv5 target detection algorithm designed in this paper has a good warning effect under various test environments.