Improving the Accuracy of 2D On-Road Object Detection Based on Deep Learning Techniques

Sammanfattning: This paper focuses on improving the accuracy of detecting on-road objects, includingcars, trucks, pedestrians, and cyclists. To meet the requirements of theembedded vision system and maintain a high speed of detection in the advanceddriving assistance system (ADAS) domain, the neural network model is designedbased on single channel images as input from a monocular camera.In the past few decades, forward collision avoidance system, a sub-system ofADAS, has been widely adopted in vehicular safety systems for its great contributionin reducing accidents. Deep neural networks, as the the-state-of-art objectdetection techniques, can be achieved in this embedded vision system withefficient computation on FPGA and high inference speed. Aimed at detectingon-road objects at a high accuracy, this paper applies an advanced end-to-endneural network, single-shot multi-box detector (SSD).In this thesis work, several experiments are carried out on how to enhance theaccuracy performance of SSD models with grayscale input. By adding properextra default boxes in high-layer feature maps and adjust the entire scale range,the detection AP over all classes has been efficiently improved around 20%, withthe mAP of SSD300 model increased from 45.1% to initially 76.8% and the mAPof SSD512 model increased from 58.5% to 78.8% on KITTI dataset. Besides,it has been verified that without color information, the model performance willnot degrade in both speed and accuracy. Experimental results were evaluatedusing Nvidia Tesla P100 GPU on KITTI Vision Benchmark Suite, Udacity annotateddataset and a short video recorded on one street in Stockholm.