cnmy-ro / Enhanced-DeepSORT

Deep-SORT object tracking algorithm with a light-weight object detection pipeline, an extension to track vehicles, and improved speed

Enhanced Deep-SORT for Vehicle Tracking

Objectives

1. Create an object detection pipeline (based on MobileNet-SSD) and integrate it with Deep-SORT. Provide two modes of operation: Eval mode for benchmarking; Cam mode for deployment. Evaluate quantitatively on MOT16, compare with original.
2. Adapt Deep-SORT to track vehicles - requires changing the CNN encoder model. Try to integrate an existing vehicle-specific model into the code. Then build and train my own model and use that instead. Evaluate qualitatively.
3. TODO -- Implement Confidence Trigger Detection mechanism for speed-up. Measure the improvement.

Useful links

Main Data set

• MOT16 Multiple Object (Pedestrian) Tracking benchmark: page | paper
• CLEAR MOT metrics: paper
• Py-MOT-metrics evaluation library: repo
• UA-DETRAC Vehicle Tracking benchmark: page

Trackers

• DeepSORT pedestrian tracking: paper | code
• SORT multi-object tracking: paper | code

Object detector

• MobileNet-SSD TF object detection API | Model zoo

Enhancement for real-time performance

• Confidence Trigger Detection paper

For vehicle tracking

• Deep-SORT for vehicle tracking code