Research On Visual Environment Perception System Under Complex Traffic Environment For Automatic Driving

Traffic road environment perception plays an important role in automatic driving and advanced driver assistance system(ADAS),which is the basis of the subsequent vehicle control and decision making.Accurate and fast perception of the traffic scenes can improve the safety of automatic vehicles.Vision-based sensors have the advantage of large information and low price when compare with other sensors,such as LIDAR and RADAR.In addition,with the rapid development of deep learning technology and the substantial improvement of hardware performance,vision-based environment perception technology is widely applied in automatic driving systems.However,the actual road traffic environment is complex and diverse,cars,pedestrians and other objects have a large difference.What's more,vision sensors are susceptible to be interfered by lighting conditions.Therefore,vision-based environment perception technology still faces great challenges.Moreover,with the limited computing power and storage space of the vehicle-mounted embedded device,the computational complexity and memory usage of perception algorithms should be small.The ”accuracy” and ”speed” of the traffic environment perception of automatic driving are concerned,and the key issues of visual environment perception around various traffic elements under complex environments are studied in this thesis.The main research contents and innovations of the thesis are summarized as follows:(1)To obtain accurate lane position information,a deeply supervised fully convolutional lane detection network is proposed.The ”hole” convolution layer is introduced to reduce the computation cost of backbone network and improve the expression ability of network.The new feature fusion module is proposed to effectively fuse the features of different layers.In order to dig into the features of each layer without increasing any test time,the deep supervision mechanism is used to supervise different output layers.This method can output accurate lane detection results at the detection speed of 25 FPS.(2)Without region proposal network and feature resampling steps,the regressionbased object detection algorithms acquire fast detection speed with low accuracy when compare with region-based detectors.For this reason,a Z-style residual object detection network(ZRNet)is proposed.The Z-style structure is aim to effectively fuse the shallow and deep feature information to obtain robust feature representation,and the location and classification accuracy is gradually improved by the residual network.ZRNet achieves a better tradeoff between object detection accuracy and speed,which significantly improves the performance of regression-based object detector.In addition,in order to solve the problems of missed detection and object jump caused by complex background,occlusion and other factors,a continuous information fusion strategy is proposed to improve detection performance and obtain more accurate and smooth detection results.(3)To optimize object detection network,the proposed residual regression model is deeply explored again,and the recurrent pyramid object detection network is proposed.Recurrent pyramid features are used during the training phase,but only the backbone detection network will be performed when testing.As the detection part is compressed,the detection speed is further accelerated.Aiming at the optimization of lane detection network,self-knowledge distillation is proposed to force the student network to learn”dark knowledge” from the complex teacher network,so the complex detection network is compressed and the operating speed is increased.In order to solve the problems of poor real-time performance and large memory consumption caused by performing object detection network and lane detection network independently,lane detection network and object detection network are integrated into a multi-task network.(4)As for the problem of poor generalization ability of the trained model due to the insufficient data and the lack of diversity,a cascade pyramid generative adversarial network is proposed to generate more samples.This method uses a cascade pyramid generator to gradually increase the resolution of generated images,a multi-scale discriminator is used to assist the generator to enrich the detailed information,and the feature matching loss devotes to stabilize training.As a way of data augmentation,the proposed cascade pyramid generative adversarial network can improve the generalization ability of the model without increasing any test time.The above-mentioned algorithms are verified on the corresponding data sets.The experimental results show that all algorithms proposed in this thesis can provide reasonable and effective solutions for traffic road environment perception.The proposed algorithms can meet the requirements of accuracy and speed,and provide a new idea for environment perception for automatic driving.