一种基于多传感器融合辅助的AlexNet模型图像识别算法

An image recognition algorithm based on multi-sensor data fusion assisted AlexNet model

自动驾驶车辆感知系统中,基于视觉的交通标志图像识别是一项关键技术。然而,由于目前硬件计算资源限制、极端光照或其他光源干扰等因素影响,在没有先验知识辅助的前提下,难以实现交通标志的准确识别和工程化应用。针对这一问题,提出了一种基于多传感器数据融合辅助的AlexNet(MSDF-AleNex)模型图像精确识别方法,利用组合导航数据对图像进行预划分,提高图像识别精度。首先,对高精度惯性/卫星导航设备、视觉传感器进行联合标定,结合高精度地图信息,获得相机与交通标志的相对位置和姿态关系;然后,利用视觉传感器自身参数,计算得到在当前图像中交通标志对应的相对位置,并据此获得224*224像素大小的目标区域;将传感器数据融合得到的目标信息和AlexNet模型相结合,目标区域作为AlexNet模型的输入数据。为验证MSDF-AlexNet模型的识别性能,基于VIVA交通信号灯数据库对模型进行离线训练并生成训练模型,然后将训练模型应用于无人巡逻车获得的实际场景交通标志图像的在线识别。结果表明,相对于AlexNet模型,MSDF-AlexNet模型在正常光照、其他光源干扰和极端光照下的综合识别精度分别达到98.4%、98%和96.8%,有助于推动系统的工程化应用。

Vision-based traffic sign image recognition is a key technology in autonomous vehicle perception system.However,due to the limitations of current hardware computing resource,extreme illumination or other light source interference,it is difficult to achieve the accurate recognition and engineering application of traffic signs without prior knowledge.To solve the problem,an accurate image recognition method based on multi-sensor data fusion-AlexNet(MSDF-AlexNet)model is proposed,which uses the combined navigation data to pre-partition the image to improve the accuracy of image recognition.Firstly,the high-precision inertial/satellite navigation equipment and visual sensor are calibrated together to obtain the relative position and attitude relationship between the camera and traffic signs by combining high-precision map information.Then,the relative position of traffic signs in the current image is calculated by using the parameters of visual sensor itself,and the target area with the size of 224*224 pixels is obtained accordingly.Combining the target information obtained by the fusion of the sensors data with the AlexNet model,the target area is used as the input data of the AlexNet model.In order to verify the recognition performance of the MSDF-AlexNet model,the traffic signal database based on the VIVA is used to train the model offline and generate training model,which is applied to the online identification of the actual traffic sign images obtained by unmanned patrol vehicles.The experiment results show that compared with the AlexNet model,the recognition accuracy of the MSDF-AlexNet model under normal light,interference from other light sources and extreme illumination is achieved 98.4%,98%and 96.8%,respectively,which is helpful to promote the engineering application of the system.