Real Time object detection and localization: Autonomous vehicles

Obstacle detection plays a major role in autonomous vehicles which has to be performed at high speed with good accuracy with constraints in computational resources. To solve this we choose Mobile Net_SSD neural network and performed experimentations to increase its performance. We initially carried out testing of MobileNet_SSD with standard convolution networks like YOLO, Faster R-CNN and Fast R-CNN. Therefore to improve the performance of MobileNet_SSD in terms of speed and accuracy we choose optimal values of its parameters such as image input resolutions, aspect ratio. We carried out tests on different versions of MobileNet (v1 & v2) and SSD on COCO and ImageNet Datasets. Later we considered different test cases to evaluate MobileNet_SSD. Finally, we demonstrate the factors responsible for best tradeoff between speed and accuracy. PASCAL VOC Dataset was trained on SSD300 and SSD512 architectures to validate the SSD framework. We successfully performed object detection on test cases such as in-motion frames, multiple-objects. For smaller input image size SSD outperforms other single stage methods in terms of accuracy.