基于改进R-CNN-SSD的智能驾驶汽车环境感知系统设计

Design of Intelligent Driving Vehicle Environment Perception System Based on Improved R-CNN-SSD

摘要:

为确保智能驾驶汽车在不同驾驶环境下均可安全行驶,提升汽车对周边环境的动态认知能力,本文设计结合改进R-CNN与SSD的智能驾驶汽车环境感知系统。系统硬件层以多源传感器为核心,感知智能驾驶汽车周围环境信息,并计算智能驾驶汽车与周边环境各个障碍物的安全距离。以该距离为依据,在软件层的支撑下,算法层结合以R-CNN模块为核心框架的两阶段目标检测算法与SSD进行智能驾驶汽车环境感知。在此过程中,利用多尺度候选框生成策略,以达到适应不同场景并提升检测能力的目的；同时引入多尺度自注意力与动态权重分配机制,捕捉到更丰富的特征关系；并在前馈传播网络(FFN)中采用非线性激活函数,缓解梯度消失问题；且采用自适应池化策略,更好地保留目标的主要特征；利用多步解码策略,逐步细化危险目标特征。最终以危险目标特征为输入,利用SSD生成汽车行驶途中小于汽车安全距离的预警框,以此完成智能驾驶汽车周边危险环境感知。实验表明:该系统可以计算与周边障碍物的安全行驶距离,并可完成不同天气条件下的驾驶环境感知,且使用该系统感知汽车周边环境后,汽车危险指数函数均小于0.12。

Abstract:

To ensure the safe operation of intelligent driving vehicles in different driving environments and enhance their dynamic cognitive ability towards the surrounding environment, this paper designs an intelligent driving vehicle environment perception system that combines improved R-CNN and SSD. The hardware layer of the system is based on multi-source sensors, which perceive the surrounding environmental information of the intelligent driving car and calculate the safe distance between the intelligent driving car and various obstacles in the surrounding environment. Based on this distance, with the support of the software layer, the algorithm layer combines a two-stage object detection algorithm with R-CNN module as the core framework and SSD for intelligent driving car environment perception. During this process, a multi-scale candidate box generation strategy is utilized to adapt to different scenarios and enhance detection capabilities; Simultaneously introducing multi-scale self attention and dynamic weight allocation mechanism to capture richer feature relationships; And adopt nonlinear activation functions in the feedforward propagation network (FFN) to alleviate the problem of gradient vanishing; And adopt an adaptive pooling strategy to better preserve the main features of the target; Using a multi-step decoding strategy to gradually refine the characteristics of dangerous targets. Finally, taking the characteristics of dangerous targets as input, SSD is used to generate warning boxes that are less than the safe distance of the car during driving, in order to achieve the perception of dangerous environments around intelligent driving cars. The experiment shows that the system can calculate the safe driving distance with surrounding obstacles, and can perceive the driving environment under different weather conditions. After using the system to perceive the surrounding environment of the car, the car hazard index function is less than 0.12.