Hardware Efficient Modified CNN Architecture for Traffic Sign Detection and Recognition

Traffic sign recognition is a vital part for any driver assistance system which can help in making complex driving decision based on the detected traffic signs. Traffic sign detection (TSD) is essential in adverse weather conditions or when the vehicle is being driven on the hilly roads. Traffic sign recognition is a complex computer vision problem as generally the signs occupy a very small portion of the entire image. A lot of research is going on to solve this issue accurately but still it has not been solved till the satisfactory performance. The goal of this paper is to propose a deep learning architecture which can be deployed on embedded platforms for driver assistant system with limited memory and computing resources without sacrificing on detection accuracy. The architecture uses various architectural modification to the well-known Convolutional Neural Network (CNN) architecture for object detection. It uses a trainable Color Transformer Network (CTN) with the existing CNN architecture for making the system invariant to illumination and light changes. The architecture uses feature fusion module for detecting small traffic signs accurately. In the proposed work, receptive field calculation is used for choosing the number of convolutional layer for prediction and the right scales for default bounding boxes. The architecture is deployed on Jetson Nano GPU Embedded development board for performance evaluation at the edge and it has been tested on well-known German Traffic Sign Detection Benchmark (GTSDB) and Tsinghua-Tencent 100k dataset. The architecture only requires 11 MB for storage which is almost ten times better than the previous architectures. The architecture has one sixth parameters than the best performing architecture and 50 times less floating point operations per second (FLOPs). The architecture achieves running time of 220[Formula: see text]ms on desktop GPU and 578 ms on Jetson Nano which is also better compared to other similar implementation. It also achieves comparable accuracy in terms of mean average precision (mAP) for both the datasets.