Total Recall: Understanding Traffic Signs Using Deep Convolutional Neural Network

Abstract

Recognizing Traffic Signs using intelligent systems can drastically reduce the number of accidents happening worldwide. With the arrival of Self-driving cars it has become a staple challenge to solve the automatic recognition of Traffic and Hand-held signs in the major streets. Various machine learning techniques like Random Forest, SVM as well as deep learning models has been proposed for classifying traffic signs. Though they reach state-of-the-art performance on a particular dataset, but fall short of tackling multiple Traffic Sign Recognition benchmarks. In this paper, we propose a novel and one-for-all architecture that aces multiple benchmarks with better overall score than the state-of-the-art architectures. Our model is made of residual convolutional blocks with hierarchical dilated skip connections joined in steps. With this we score 99.33% Accuracy in German sign recognition benchmark and 99.17% Accuracy in Belgian traffic sign classification benchmark. Moreover, we propose a newly devised dilated residual learning representation technique which is very low in both memory and computational complexity.

Traffic Sign Detection and Recognition Using Multi-Scale Fusion and Prime Sample Attention

“…Traditional object detection algorithms are still widely used for fast calculation speed and low memory footprint [10][11][12][13] . For example, Anant Ram Dubey [14] et al used HOG-SVM method to detect road objects, and Takaki Masanari [15] used SIFT [6] method to detect traffic signs.…”

Robust Semi-Supervised Traffic Sign Recognition via Self-Training and Weakly-Supervised Learning

“…The result in Table 3 provides a comparison of the classification accuracy between ROSST, the proposed method, and some state-of-the-art supervised learning algorithms, which include single CNN with three STNs [29], DCGAN-PILAE [30], traffic sign classification based on pLSA [16], multiscale CNNs [31], BAGAN [32], traffic sign recognition with hinge loss CNNs (HLSGD) [49], and residual blocks CNN [50]. All these state-of-the-art methods were evaluated on the GTSRB dataset, making it fair to compare the proposed method ROSST with them.…”

“…This occurred because those frameworks had accuracy rates that were higher than the proposed method. The state-of-the-art models [29,30,49,50] used all of the training set to train the model and then were tested on the test set, but the proposed method used 60% of the training data to train and validate the model during the supervised training stage and later ran on the unlabeled set(the remainder of training plus the test sets) for the self-paced learning phase. This showed that the model could perform well and achieve a higher recognition rate on a small set of data and train a good model just as the state-of-the-art models.…”

“…Single CNN with 3 STNs [29] 99.71 DCGAN-PILAE [30] 99.80 pLSA [16] 98.14 Multiscale CNNs [31] 98.84 BAGAN [32] 96.75 HLSGD [49] 99.65 Residual blocks CNN [50] 99.33 ROSST (our approach) 99.27…”

“…Residual blocks CNN [50] 99.17 Single CNN with 3 STNs [29] 99.71 DCGAN-PILAE [30] 99.80 VGG-16 [51] 99.72 ROSST (our approach)…”