基于RCNN的车辆检测方法研究

Algorithm of vehicle detection based on RCNN

为了解决基于传统机器学习车辆检测算法实时性和泛化能力差的问题,对基于深度学习的车辆检测算法进行研究。分析了Faster R-CNN检测算法原理,使用Python编程语言基于深度学习框架Tensor Flow实现了Faster R-CNN算法;采集了4个季节不同天气情况下的12 000张路况图片数据,并对数据集进行了标注,采用3种不同方式对数据集进行了预处理;通过对照试验对Faster RCNN算法中的超参数进行了调优;使用控制变量法比较了RCNN、SPPnet、Fast R-CNN和Faster R-CNN 4种算法的检测准确率和检测速度,指出了4种算法的主要耗时步骤,验证了Faster R-CNN车辆检测算法的有效性。研究结果表明:基于Faster R-CNN的车辆检测算法达到每张69 ms的检测速度和91.3%的准确率,能够实现实时高精度的车辆检测。

Aiming at improving the robust adaptation and real-time of the vehicle detection algorithm based on traditional machine learning,vehicle detection algorithm based on deep learning was researched. The principle of Faster R-CNN detection algorithm was analyzed. Based on Tensor Flow deep learning frame and made use of Python programming language,the Faster R-CNN algorithm was realized. The road condition data set of four seasons was collected and labeled,12 000 pictures was included. The data set was pretreat by three different method and the algorithm parameter of Faster R-CNN was tuned by controlled experiment. The accuracy and speed of four detection algorithm,RCNN,SPPnet,Fast R-CNN and Faster R-CNN,was compared by variable-controlling approach. The results indicate that the speed of vehicle detection based on Faster R-CNN is 69 ms and the accuracy rate is 91. 3%,the algorithm can realize real-time and high accuracy of vehicle detection.