Multi-object Tracking using Deep Learning

Multi-object Tracking (MOT) is a technique to identify or detect and track objects in images and videos. We introduce a deep learning model with just one complex neural network. The method, detection by tracking uses a detector called SSD (Single Shot Detector) a deep learning network which is trained and evaluated on the Pascal VOC (2007+2012) dataset based on 20 different object classes which has over 16k images used for training and over 4k images used for evaluation of the detection model. SSD discretizes the outcome area by creating prior boxes having varying aspect ratios and scales of feature maps and arranging them into bounding boxes for every position. Moreover, the detection output integrates with the tracker from various feature maps with varying resolutions to effectively manage and track the moving objects of different dimensions in a video. Results from experimentation are conducted on the PASCAL VOOC2007 test dataset, SSD attains a 73% mAP for input frame of size $300\times 300$. The accuracy is similar to progressive model. In comparison, the performance is further enhanced with the addition of tracker to other single stage techniques.