Object Detection System Based On Deep Convolutional Neural Network

Object detection is a classic problem in the field of multimedia and computer vision. In recent years, the rapid development of deep learning has greatly promoted the object deteetion research. In this paper, we implement an object detection algorithm based on deep convolutional neural network combined with regression method. Based on this algorithm, we develop a real-time object detection system on the low-power embedded platform NVIDIA Jetson TX1.The past algorithms such as R-CNN, Fast R-CNN and Faster R-CNN treat object detection as a classification task after extracting feature in the region proposal. These models can get high precision rate, but can not achieve real-time speed. Our algorithm takes object detection as a regression task, constructs a multi-task loss function to combine the regression of the boundingbox and the object classification. Achieved 120 FPS on the computer with GPU acceleration.Because of the limited memeory of CPU and GPU on the Jetson TX1, our model balances the depth and the size. Algorithm consists of eight convolutional layers and one detection layer. In the convolutional layers, take 3×3 kernels and followed by Batch Normalization and MaxPooling. The algorithm reduces the size by using mostly convolutional layers without the large fully connected layers at the end.Algorithm training consists of pre-training and fine-tuning, training the deep convolutional neural network on the ILSVRC2012 dataset, achieved top-1 and top-5 accuracy rate of 58.3% and 81.3%. Do the fine-tuning on the Pascal VOC2007 and VOC2012, achieved 54.2 mAP on VOC 2007 validation and 47.8 mAP on VOC2012 validation.After the training on the general x86 computer with GPU acceleration,transplanted the algorithm to the embedded platform NVIDIA Jetson TX1.Based on this algorithm, develop the real-time object detection system which consists of the camera input, image frame pre-processing, object detection algorithm and the video output preprocessing. The system achieved 26 FPS with power consumption only 11W on Jetson TX1.