Research And Application On Object Detection Algorithms For Autonomous Driving Scenarios

As the main development direction of intelligent and interconnected transportation,autonomous driving technology can provide security for traffic and improve transportation efficiency,thereby bringing convenience for people.Computer vision plays a vital role in autonomous driving.As a crucial component of computer vision,object detection is definitely a critical solution to autonomous driving.In recent years,deep learning technology has set off another wave of technology,updating kernel technologies in computer vision,and providing great assistance to the real landing of autonomous driving.However,due to the impact of various vehicle scales on the road,complex and changeable scenes,lighting changes,and occlusion,the performance of object detection algorithms for autonomous driving scenarios needs to be improved.Against this backdrop,this thesis has conducted in-depth research on object detection algorithms for autonomous driving scenarios.The main research contents include the following aspects:(1)Object detection for autonomous driving scenarios requires concerns on accuracy,speed and sensitivity to multi-scale objects,especially small objects.Aiming at this problem and the disadvantage of one-stage algorithm in processing foreground-background class imbalance encountered during training,this thesis proposes BANet,which combines the concept of atrous convolution and feature fusion to improve the network structure of the original SSD thereby expanding receptive field and enriching semantic information in shallow layers.Meanwhile,a new loss which handles the class imbalance is designed to replace the standard cross entropy loss in the original algorithm.The experimental results indicate that BANet surpasses the original SSD in m AP on KITTI with a relatively high speed retained.(2)In order to reduce the storage space and calculation amount of the model and reduce its dependence on hardware configuration,this thesis proposes Tiny FCOS,a lightweight single-stage object detection algorithm,which absorbs FCOS's advantages,that is,avoiding all the complex calculations and hyperparameters related to anchor,while applying its pixel-based structure in combination with the techniques in semantic segmentation so as to further explore its potential.The backbone is constructed by the basic components from LEDNet,which is a lightweight semantic segmentation algorithm.And the feature pyramid network is built by standardized hybrid dilated convolution in order to simplify the prediction branches of FCOS.Tiny FCOS outperforms Tiny YOLOv3 in m AP on dataset PASCAL VOC,while shares similar amount of parameters and calculation,and speed with the latter.Meanwhile,the experimental results on dataset KITTI also indicates that Tiny FCOS is also competent for the high-resolution input of autonomous driving scenario.(3)The algorithms proposed above are applied to the object detection system for the automatic driving scenario thereby realizing real-time detection and recognition of vechicles and pedestrians on the road and visual representation,which is implemented by Python GUI.