Font Size: a A A

Research On Object Detection And Tracking Algorithms For Intelligent Vehicles In Complex Scenarios

Posted on:2022-12-24Degree:DoctorType:Dissertation
Country:ChinaCandidate:J W CaoFull Text:PDF
GTID:1482306758477354Subject:Vehicle Engineering
Abstract/Summary:PDF Full Text Request
With the increase of car ownership year by year,urban traffic congestion is becoming more and more serious,and road traffic accidents occur frequently.Intelligent vehicles can assist or even completely replace the driver to complete the main driving manipulation based on real-time road conditions,gradually becoming an ideal solution to traffic safety issues.Environmental perception is a vital part of the intelligent vehicle integrated system,which is equivalent to the "eyes" of the vehicle.It directly affects the intelligent level of decision-making system and control system,and is a prerequisite for the safe driving of autonomous vehicles.Compared with other on-board sensors,the vision sensor based on camera can obtain the richest environmental information,with low cost,mature hardware technology,so it has unique advantages in the field of autonomous driving.In recent years,due to the great success of deep learning methods in the field of image processing and pattern recognition,emerging technologies represented by computer vision have pointed out a new development direction for the research of intelligent vehicle visual perception.Although the existing visual perception algorithms have achieved some fruitful research results,they can only be applied to simple environments with good scenes.Many factors such as insufficient light,bad weather conditions,occlusion of obstacles,and cluttered backgrounds have brought a series of challenges and trials to the long-term development of visual perception algorithms.Facing the complex and changeable external environments,how to realize the rapid processing and accurate understanding of visual information is a problem that needs to be solved urgently in the current research on the visual perception of smart cars.In actual road scenes,the main target objects existing in the surrounding environment of the vehicle include lane lines,traffic signs,and obstacles such as pedestrians and vehicles.Therefore,this article is committed to in-depth research on the object detection and tracking algorithms for intelligent vehicles in complex scenarios.The specific research contents are as follows:(1)A lane detection algorithm for complex road conditions and dynamic environments is proposed.The distorted image is converted by camera calibration and image distortion removal,the edge detection algorithm of superimposed threshold is adopted to determine the basic contour of the lane line,and the region of interest(ROI)extraction and inverse perspective transformation are used to obtain a bird's-eye view of the lane line.Considering that the B-spline curve has the properties of affine invariance,local modification and strong convex hull,it can flexibly express the shape of the curve.Based on the third-order B-spline curve model,the random sample consensus(RANSAC)algorithm is used to curve-fit the lane lines,and fitting evaluation and curvature radius calculation are then carried out on the curve.The testing experiment results show that the average accurate recognition rate of the proposed algorithm based on road driving videos reaches 98.49%,and the average processing time per frame is 21.5ms.The average accurate recognition rate based on Tu Simple dataset reaches 98.42%,and the average processing time of each frame is 22.2ms.(2)A traffic sign detection and recognition algorithm based on convolutional neural network is proposed.The traffic signs are segmented based on HSV color space,and the binary image is morphologically processed to achieve effective extraction of traffic signs.The classic LeNet-5 convolutional neural network model is only suitable for the classification and recognition of single object,so we improved it greatly,including using Gabor kernel as the initial convolution kernel,adding the batch normalization(BN)processing after the pooling layer,selecting Adam method as the optimizer algorithm and adding Dropout in the fully-connected layer.The experimental results indicate that the improved LeNet-5 network model has good generalization ability,and it can classify and recognize different types of traffic signs accurately and efficiently.The accurate recognition rate of the proposed algorithm based on GTSRB dataset reaches 99.75%,and the average processing time per frame is 5.4 ms.(3)A pedestrian detection algorithm based on the improved YOLOv3 model is proposed.Considering that pedestrians are typical non-rigid objects,the poses between pedestrians are different.When using the YOLOv3 network model to detect pedestrians,it is prone to inaccurate object positioning,missed detection of small-scale objects,and low detection accuracy in object-intensive scenes,reasonable improvements are made.In the improved network model,the division size of grid cells is adjusted,the improved k-means clustering algorithm is adopted,the multi-scale bounding box prediction based on expanded receptive field is used,and the Soft-NMS algorithm is employed.The experimental results show that the m AP value of the proposed algorithm based on INRIA person dataset reaches90.42%,which is 6.88% higher than the detection accuracy before the improvement.The average processing time of each frame is 9.6ms,which is 3.9ms shorter than the detection speed before the improvement.The m AP value based on the PASCAL VOC 2012 dataset reaches 91.14%,which is 7.14% higher than the detection accuracy before the improvement.(4)A vehicle detection algorithm based on the improved SSD model is proposed.Considering that vehicles are typical rigid objects,the similarity between vehicles is high.When using the SSD network model to detect vehicles,it is easy to have insufficient feature extraction ability of small-scale objects,easy missed detection of occluded objects and poor detection performance in complex environments,targeted improvements are made.In the improved network model,the basic structure of the network is redesigned by combining the inception blocks and feature fusion methods,the weighted mask is used in network training,and the repulsion loss is added to the loss function.The experimental results show that the m AP value of the proposed algorithm based on KITTI dataset reaches 92.18%,which is4.64% higher than the detection accuracy before the improvement.The average processing time of each frame is 15 ms,which is 13 ms shorter than the detection speed before the improvement.The average m AP value based on the self-made vehicle dataset reaches91.76%,and the detection accuracy under the same weather condition is higher than that before the improvement of the network model.(5)An object tracking algorithm based on double-template Siamese network model is proposed.Aiming at the problems of existing Siamese network algorithms in object tracking,such as insufficient ability to extract deep features,ignoring the detailed information of shallow features,and unable to ensure the balance between tracking accuracy and tracking speed,a double-template Siamese network model based on multi-feature fusion is constructed on the basis of Siam-FC network.In the designed network model,the improved lightweight network Mobile Net V2 is used as the feature extraction backbone network,the attention mechanism is introduced to realize the focus on the key details of the object,and the template online update mechanism is introduced to deal with the problem of occlusion of the object.The experimental results show that the average overlap rate of the proposed algorithm based on the OTB2015 dataset reaches 82.5%,the average center location error is8.5 pixels,and the tracking speed is 56 FPS.In the actual driving video sequences,the average overlap rate reaches 84.7%,the average center location error is 6.9 pixels,and the tracking speed is 58 FPS.It can be seen from the test and evaluation results that the proposed object detection and tracking algorithms have good accuracy and real-time performance,and can still maintain strong robustness and anti-interference ability in complex scenarios,effectively solving the problems and deficiencies of existing research methods.The research contents in this paper have important theoretical significance and engineering application value,and are conducive to promoting the further development and improvement of the technical level of intelligent vehicle assisted driving.
Keywords/Search Tags:Intelligent Vehicles, Environmental Perception, Computer Vision, Deep Learning, Object Detection, Object Tracking
PDF Full Text Request
Related items