Traffic sign detection with VG-RAM weightless neural networks

Abstract

We present a biologically inspired approach to traffic sign detection based on Virtual Generalizing Random Access Memory Weightless Neural Networks (VG-RAM WNN). VG-RAM WNN are effective machine learning tools that offer simple implementation and fast training and test. Our VG-RAM WNN architecture models the saccadic eye movement system and the transformations suffered by the images captured by the eyes from the retina to the superior colliculus in the mammalian brain. We evaluated the performance of our VG-RAM WNN system on traffic sign detection using the German Traffic Sign Detection Benchmark (GTSDB). Using only 12 traffic sign images for training, our system was ranked between the first 16 methods for the prohibitory category in the German Traffic Sign Detection Competition, part of the IJCNN'2013. Our experimental results showed that our approach is capable of reliably and efficiently detect a large variety of traffic sign categories using a few training samples.

Occupancy Grid Mapping with Cognitive Plausibility for Autonomous Driving Applications

“…The saccadic system has inspired similar mechanisms in computer vision. For example, De Souza et al [7] implement a variant of the retinal log-polar transform for a neural network to detect traffic signs using a small dataset. Note that there are also computational models that implement aspects of natural visual attention accurately [8,9,19,21].…”

Self-Driving Cars: A Survey

“…Furthermore, some methods worked only with recognition and not with detection, perhaps because of the lack of data. Only after large databases were made available (such as the well-known German Traffic Sign Recognition (GTSRB) (Stallkamp et al, 2012) and Detection (GTSDB) (Houben et al, 2013) Benchmarks, with 51,839 and 900 frames, respectively) that learning-based approaches (Houben et al, 2013;Mathias et al, 2013) could finally show their power, although some of them were able to cope with fewer examples (De Souza et al, 2013a). With the release of even larger databases (such as STSD (Larsson & Felsberg, 2011) with over 20,000 frames, LISA (Jensen et al, 2016b) with 6,610 frames, BTS (Mathias et al, 2013) 25,634 frames for detection and 7,125 frames for classification, and Tsinghua-Tencent 100K (Zhu et al, 2016) with 100,000 frames), learning-based approaches improved and achieved far better results when compared to their model-based counterparts.…”

“…It is part of the TSD subsystem. The Traffic Sign Detector module (De Souza et al, 2013b;Torres et al, 2019) detects and recognizes (Berger et al, 2013) traffic signs along the path from images captured by the front camera (see video at https://youtu.be/SZ9w1XBWJqE). It is part of the TSD subsystem.…”

Traffic sign detection in MLS acquired point clouds for geometric and image-based semantic inventory

“…Once an entire cluster is projected into an image, the bounding box of the projected points is computed, and the original image is cropped with a margin of 25% of the bounding box size, to deal with calibration errors and add some background to the detection (Figure 5d). This way, several images of the same sign can be obtained, and the recognition process can be carried out using RGB images with the following advantages with respect to approaches that only use images (Soheilian et al 2013, De Souza et al 2013): 1) The detection problem is almost solved, as the location of the traffic signs is known beforehand. 2) Images of the same traffic sign can be stored and analysed together, hence a single classification result is expected for each set of input images ( Figure 6).…”