Neural network is one of the most widely used method in autonomous driving. Current researchers use only steering angle to train artificial neural networks, ignoring the importance of acceleration and deceleration for autonomous vehicles. We used an intelligent driving platform built with the Raspberry Pi 4 Model B, a front wide-angle camera, and a 1:16 scale model car to achieve real-time acceleration and deceleration while performing road tracking. Existing models cannot learn steering angle and throttle values well. This research proposed a novel architecture CNN model (PBLM-CNN21) to achieve real-time acceleration and deceleration while achieving road tracking. The PBLM-CNN21 model can learn steering angle and throttle value. The training loss value of our proposed PBLM-CNN21 model was 35% lower than the current TDD model, and the stability of our proposed road tracking model was 82% greater than that of the current TDD model. Furthermore, we tested the impact of different hyper-parameters on training model loss and road tracking performance. In addition, we also tested the effectiveness of varying lighting conditions and speed ratios on road tracking performance. The PBLM-CNN21 model proved more robust than the existing TDD models. Moreover, the PBLM-CNN21 model achieved road tracking under different lighting conditions and was more suitable for high-speed ratios. Keywords: autonomous driving; raspberry pi; deep learning; road tracking; convolutional neural networks; PBLM-CNN21 *Corresponding author: Tel.: (+66) 0966033973 E-mail: jianqu@pim.ac.th
Purpose: The purpose of this research is to achieve multi-task autonomous driving by adjusting the network architecture of the model. Meanwhile, after achieving multi-task autonomous driving, the authors found that the trained neural network model performs poorly in untrained scenarios. Therefore, the authors proposed to improve the transfer efficiency of the model for new scenarios through transfer learning.Design/methodology/approach: First, the authors achieved multi-task autonomous driving by training a model combining convolutional neural network and different structured long short-term memory (LSTM) layers. Second, the authors achieved fast transfer of neural network models in new scenarios by cross-model transfer learning. Finally, the authors combined data collection and data labeling to improve the efficiency of deep learning. Furthermore, the authors verified that the model has good robustness through light and shadow test.Findings: This research achieved road tracking, real-time acceleration–deceleration, obstacle avoidance and left/right sign recognition. The model proposed by the authors (UniBiCLSTM) outperforms the existing models tested with model cars in terms of autonomous driving performance. Furthermore, the CMTL-UniBiCL-RL model trained by the authors through cross-model transfer learning improves the efficiency of model adaptation to new scenarios. Meanwhile, this research proposed an automatic data annotation method, which can save 1/4 of the time for deep learning.Originality/value: This research provided novel solutions in the achievement of multi-task autonomous driving and neural network model scenario for transfer learning. The experiment was achieved on a single camera with an embedded chip and a scale model car, which is expected to simplify the hardware for autonomous driving.
19 April 2023 in CURRENT APPLIED SCIENCE AND TECHNOLOGY
Despite the prevalence of scams in crowdfunding projects, currently, there is little research into the identification of fraudulent or infeasible crowdfunding projects. Since detecting fraudulent crowdfunding projects is challenging, most existing research on fake information has focused on detecting fake news or fake charity crowdfunding projects based on social media, but research on fraudulent or infeasible crowdfunding projects is very limited. Therefore, to solve this problem, we focus on how to detect fraudulent crowdfunding projects based on knowledge extraction and contradiction detection. We proposed a novel method called BM5-SP-SC (BERT-MT5-Sentence Pattern-Sentiment Classification). BM5 (BERT-MT5), which is built from a combination of a key-BERT and a fine-tuned MT5 transformers, was used to extract feature information from crowdfunding projects. We proposed a novel method for MT5 training to construct an adaptive BM5 model. The correct rate of keywords extracted by our novel adaptive BM5 model was up to 72.7%, the recall was 100%, and the F-measure was up to 84.19%. The minimum train loss of the BM5 model was up to 0.1342, and the evaluation loss achieved was 0.3064. The BLEU score of summary-to-keyword was 37.336. Moreover, we proposed an SP (Sentence Pattern) matching method to achieve knowledge extraction. Furthermore, SC (Sentiment Classification) was used to build a sentiment classifier thesaurus for identifying fraudulent and infeasible crowdfunding projects. Our proposed BM5-SP-SC achieved an overall accuracy of 85.26% in detecting fraudulent crowdfunding projects. Keywords: crowdfunding projects; contradiction detection; knowledge extraction; BERT; MT5; feature information extraction *Corresponding author: Tel.: (+66) 863759307 E-mail: jianqu@pim.ac.th
WaDang recognition can provide valuable information for digital protection of cultural relics. However, WaDang recognition is challenging due to a great number of variations, such as different characters, different background pattern, and different scale and rotation. In this paper, we try to solve these problems by using hand-crafted and deep features and integrate them into multiple instance learning, in which the features of instances are divided into two parts and their similarity is computed by multiple kernel learning. The experiments on our collected WaDang images, MUSK, and COREL datasets show that the proposed algorithm is effective and its performance is compared with other state-of-art algorithms.
14 December 2024 in ASEAN Journal of Scientific and Technological Reports
Adversarial attacks are a significant threat to autonomous driving safety, especially in the physical world where there is a prevalence of "sticker-paste" attacks on traffic signs. However, most of these attacks are single-category attacks with little interference effect. This paper builds an autonomous driving platform and conducts extensive experiments on five single-category attacks. Moreover, we proposed a new physical attack - a mixed attack consisting of different single-category physical attacks. The proposed method outperforms existing methods and can reduce the accuracy of traffic sign recognition of an autonomous driving platform by 38%. Furthermore, we proposed a new anti-jamming model for physical adversarial defense, CBAM-ResNet26 & CBAM-Alexnet, which improves an autonomous driving platform's traffic sign recognition accuracy to 63% under mixed attack. Finally, experiments were also conducted with datasets with different ratios of adversarial attack examples, and the experimental results showed that in adversarial training, the higher the ratio of adversarial examples, the higher the recognition accuracy. However, a too high ratio would reduce the accuracy of normal traffic signs. Finally, the optimal ratio for physical adversarial defense training is 1:2.