Research On Road Scenes Object Detection Based On Deep Learning

With the development of economy and the promulgation of encouragement policies,China’s automobile industry has been developing at a high speed.With the number of domestic car ownership increasing,traveling by car not only brings convenience to people,but also brings great security threats.In recent years,artificial intelligence has been arisen and the concept of intelligent driving has also been proposed.As the first step of intelligent driving,object detection needs to identify the targets such as vehicle,non-motor vehicle and pedestrian in front of the road accurately.But in practical application scenarios,road conditions are complex,factors which variant to the illumination,poses and truncation will affect the precision of object detection system,and then it will affect the safety of intelligent driving.Object detection algorithm based on deep learning depend on a deep Convolutional neural networks,and has a strong learning and scenario analysis ability.At present,the main object detection algorithm based on deep learning includes two-stage detection with region proposals represented by Faster-RCNN and a single-stage detection with regression represented by SSD and YOLO.Considering the high-speed requirements of the detection algorithm in the intelligent driving scene,this paper studies on a single-stage detection.The details are as follows:1.I have studied on the basic work related to object detection,and analyzed the traditional object detection algorithm and the object detection method based on deep learning,and sorted out the model structure of convolutional neural network.The geometric structure of the traditional convolutional layer is fixed,the receptive field of the convolution kernel is limited,lacking the internal change mechanism,so it cannot adapt to the scale variation and part deformation.In order to solve the problem of internal covariate shift in model training,batch normalization is often used to adjust the data distribution.However,this method only retains the difference between single samples,making the network vulnerable to appearance changes.2.For the typical algorithms YOLO and YOLOv2 in the single-stage object detection algorithm,I have studied on it’s the network structure,advantages and disadvantages of the algorithm and improved the accuracy of the algorithm by modifying the basic network of YOLOv2 algorithm.In this paper,I combined the YOLOv2 with the Resnet101 feature extraction network,which computational power and performance are suitable.The comparison on KITTI data set proves the effectiveness of the fusion method,and the precision is improved.3.Based on the YOLOv2 fusion network,this paper improves the structure of the convolutional layer and method of normalization.Modulated deformable convolution was used to replace the upper convolutional layer in the deep network,and the IBN was used for data normalization at the bottom of the network.The improved algorithm works well on KITTI dataset,improving the detection precision of non-motorized vehicles and pedestrians with large deformation objects.The algorithm of this paper is tested in the self-acquired driving video.The results show that the algorithm can adapt to the scenes with poor lighting conditions such as fog and tunnel.