Vehicles detection using deep learning with improved single shot detector

Abstract

Deep learning is an important element of data science to automate predictive analysis for a computer to detect and classify the objects into different classes based on trained datasets, either through supervised learning, semi-supervised learning, or unsupervised learning. The aim of this research work is to use a deep learning algorithm, improved single shot detector (SSD) which is capable to detect vehicles, so that the proposed algorithm not only achieve fast detection speed, but also achieve high accuracy in object detection. Although there is other algorithm available in the context of deep learning such as You Only Look Once (YOLO) and Faster-Region Based Convolutional Neural Networks (Faster R-CNN), most of them have trade off between accuracy and speed in object detection. The accuracy also degrades when detecting small objects or objects that are further away. Besides, current network models have difficulties in identifying objects by relying solely on the pre-trained datasets, as the traffic participants may vary across cities, with different colours and shapes. Furthermore, training datasets manually with a variety of car models’ images would be time consuming due to the huge datasets. Hence, one of the research objectives is to implement mobilenet V2 network architecture on existing SSD network to improve the detection accuracy (mAP, mean average precision), inference time (s, second) and sensitivity towards small objects in complex backgrounds without increasing the computation complexity. The second research objective of this project is to apply transfer learning mechanism for the custom dataset to increase detection accuracy against small objects and reduce training time. In this research, custom datasets are used for training and testing, where the datasets are annotated using labelImg. Google Colab and some open-source libraries, Tensorflow and Keras will be used in model training. The performance of improved-SSD in object detection is evaluated based on inference time (second) and mean average precision (mAP). All models are pretrained using Common Objects in Context dataset (COCO). On top of that, own custom dataset is used to archive better accuracy. Based on the result obtained, 1.76 seconds needed for Faster R-CNN model to perform inference per image whereas 1.24 seconds needed for proposed model to perform the same tasks. The inference time of proposed model is approximate 30% faster than the Faster RCNN model. The mean average precision of the proposed model is 73.4% whereas the average recall rate of the proposed model is 80%. Besides, the proposed model obtains approximately 10% improvement in terms of mAP detecting small object if compared with Faster R-CNN model. The proposed model able to detect vehicles with shorter inference time and good accuracy. The model shows improvement in detecting small objects and objects that are further away with better accuracy. In short, the trained model can serve as a good starting point for the development of autonomous car.