Project 9

DEEPSORT

DeepSORT is a realtime Object Tracking Algorithm based on Deep Learning. It is an extension of the Simple Online Realtime Tracking algorithm. The issue with the SORT algorithm is that it fails during Occlusion and there are a large number of ID switches during person re-identification. DeepSORT solves these problems by using a Deep association matrix. A convolutional neural network is used to train the matrix offline which is used specifically for person re-identification.

The Association matrix is trained by modelling an association problem with the following two metrics:

1. Mahalanobis distance
2. Cosine distance

The Association matrix is implemented as a deep neural network and the matrix is used online(real-time) during tracking(inference) to minimise the number of ID switches as well as to help with tracking objects during occlusion.

DeepSort is a tracking algorithm and apriori knowledge of the position of the object is required to perform the tracking. DeepSORT and SORT use the Kalman filter Estimation method to calculate the position of the Object. Object Detection can be performed using multiple Object detection algorithms such as:

1. R-CNN
2. Faster R-CNN
3. YOLO
4. SSD

In this particular case, YOLO v5 is used to provide the bounding box coordinates for the Objects of interest, In this particular case, Vehicles are detected using YOLO v5.

The Output video in which the tracking of the vehicles is done is attached below.
(Alternative Links - https://drive.google.com/file/d/1t-G58RZXfcHisGjCahiiRwtSd-YgIaWA/view?usp=sharing)