Lane detection plays a vital role in making the idea of the autonomous car a reality. Traditional lane detection methods need extensive hand-crafted features and post-processing techniques, which make the models specific feature-oriented, and susceptible to instability for the variations on road scenes. In recent years, Deep Learning (DL) models, especially Convolutional Neural Network (CNN) models have been proposed and utilized to perform pixel-level lane segmentation. However, most of the methods focus on achieving high accuracy while considering structured roads and good weather conditions and do not put emphasis on testing their models on defected roads, especially ones with blurry lane lines, no lane lines, and cracked pavements, which are predominant in the real world. Moreover, many of these CNN-based models have complex structures and require high-end systems to operate, which makes them quite unsuitable for being implemented in embedded devices. Considering these shortcomings, in this paper, we have introduced a novel CNN model named LLDNet based on an encoder–decoder architecture that is lightweight and has been tested in adverse weather as well as road conditions. A channel attention and spatial attention module are integrated into the designed architecture to refine the feature maps for achieving outstanding results with a lower number of parameters. We have used a hybrid dataset to train our model, which was created by combining two separate datasets, and have compared the model with a few state-of-the-art encoder–decoder architectures. Numerical results on the utilized dataset show that our model surpasses the compared methods in terms of dice coefficient, IoU, and the size of the models. Moreover, we carried out extensive experiments on the videos of different roads in Bangladesh. The visualization results exhibit that our model can detect the lanes accurately in both structured and defected roads and adverse weather conditions. Experimental results elicit that our designed method is capable of detecting lanes accurately and is ready for practical implementation.

Inception layers are most effective and constructive architecture in current object detection field. Moreover, beside Inception layers, feature fusion layers also shows significant performance for object detection. Single Shot Multibox Detector(SSD) is one of the most convenient and fastest algorithms in the current object detection field. SSD uses only a single convolutional neural network to detect the object from the input. Moreover, SSD algorithm shows high accuracy to detect multi-scale objects but detecting small objects SSD network performance is not satisfactory. In this paper, we added feature fusion layers and Inception layers with SSD to increase its performance. In this paper, our main focus to improve the features of different layers by feature fusion and also minimize the calculation cost by Inception. Moreover, here in SSD, we add extra layers to improve its performance without any affecting on its speed. By adding Inception layers and feature fusion layers, SSD network can catch more information without increasing the complexity of the network and it also able to detect more small objects. The new proposed algorithm can detect small and big objects with satisfactory accuracy.

Generative adversarial networks (GANs) achieved impressive performance on image generation and visual classification applications. However, adversarial networks meet difficulties in combining the generative model and unstable training process. To overcome the problem, we combined the deep residual network with upsampling convolutional layers to construct the generative network. Moreover, the study shows that image generation and classification performance become more prominent when the residual layers include on the generator. The proposed network empirically shows that the ability to generate images with higher visual accuracy provided certain amounts of additional complexity using proper regularization techniques. Experimental evaluation shows that the proposed method is superior to image generation and classification tasks.

실시간 물체 탐지는 현재 물체 탐지 및 분류 분야에서 가장 중요하고 도전적인 과제 중 하나이다. 탐지 및 분류 정확성을 향상시키기 위해 지난 수십 년 동안 많은 연구가 이루어졌다. 비디오 감시, 로봇 비젼, 자율주행 차랑 및 많은 어플리케이션의 정확성을 향상시키기 위해 컴퓨터 비젼에 대한 연구가 증가하고 있다. YOLO(You Only Look Once) 는 가장 빠른 CNN (Convolutional Neural Network) 중 하나이며 실시간 객체 탐지분야에 최신 기술로 자리 잡았다. 그러나 여전히 약간의 지역화 문제점이 존재한다. 본 논문에서는 YOLO Network의 성능을 향상 시키기 위해 feature filter bank를 추가한 network를 제안한다. 실험을 통해 제안하는 network의 객체 검출 성능이 향상됨을 증명하였다.

컨볼루션 필터는 기본적으로 집합 분류에 따라서 입력을 특징화한다. 이런 필터는 실제 구조와 동일하게 입력을 재구성하기 위해 단순한 토폴로지 구조를 조절한다. CNN(Convolutional Neural Networks)은 객체를 분류하기 위해 변하지 않는 필터 배열들을 사용한다. 본 논문에서는 오프셋 학습을 통한 객체를 분류하는 특유의 컨볼루션 필터 구조를 제안한다. 정확한 객체 분류를 위한 객체 국소화 성능을 향상하기 위해서 네트워크는 공간적 증가 국소화를 사용한다. 이런 국소화 절차는 오프셋 학습의 개념을 따르고 변하지 않는 필터 배열이 아닌 개선된 변형 가능한 컨볼루션 구조를 사용하여 테스트 데이터 셋을 증가시킵니다. 특징 추출 및 국소화를 위해 픽셀 정보를 분류하는 오프셋 학습을 채택한다. 또한, 프레임워크는 비용 계산을 줄이기 위해 가장 가까운 픽셀 값을 계산하는 오프셋 분류를 사용하여 객체 영역을 처리한다. 증가된 테스트 데이터 셋을 통해 더 정확한 검출 작업이 가능하다.