开源软件名称：

Deep-SORT-YOLOv4

开源软件地址：

https://gitee.com/ywuwgij/Deep-SORT-YOLOv4

开源软件介绍：

Introduction

This project was inspired by:

I swapped out YOLO v3 for YOLO v4 and added the option for asynchronous processing, which significantly improvesthe FPS. However, FPS monitoring is disabled when asynchronous processing is used since it isn't accurate.

In addition, I took the algorithm from this paper and implemented it into deep_sort/track.py.The original method for confirming tracks was based simply on the number of times an object has been detected without considering detection confidence, leading to high tracking false positive rates when unreliable detections occur (i.e. low confidence true positives or high confidence false positives). The track filtering algorithm reduces this significantly by calculating the average detection confidence over a set number of detections before confirming a track.

See the comparison video below.

Low confidence track filtering

Navigate to the appropriate folder to use low confidence track filtering. The above video demonstrates the difference.

See the settings section for parameter instructions.

YOLO v3 and YOLO v4 comparison video with Deep SORT

With asynchronous processing

As you can see in the gif, asynchronous processing has better FPS but causes stuttering.

This code only detects and tracks people, but can be changed to detect other objects by changing lines 103 in yolo.py. For example, to detect people and cars, change

if predicted_class != 'person':    continue

if predicted_class not in ('person', 'car'):    continue

Performance

Real-time FPS with video writing:

~4.3fps with YOLO v3
~10.6fps with YOLO v4

Turning off tracking gave ~12.5fps with YOLO v4.

YOLO v4 performs much faster and appears to be more stable than YOLO v3. All tests were done using an Nvidia GTX 1070 8gb GPUand an i7-8700k CPU.

Quick start

Download and convert the Darknet YOLO v4 model to a Keras model by modifying convert.py accordingly and run:

python convert.py

Then run demo.py:

python demo.py

Settings

Normal Deep SORT

By default, tracking and video writing is on and asynchronous processing is off. These can be edited in demo.py by changing:

tracking = TruewriteVideo_flag = TrueasyncVideo_flag = False

To change target file in demo.py:

file_path = 'video.webm'

To change output settings in demo.py:

out = cv2.VideoWriter('output_yolov4.avi', fourcc, 30, (w, h))

Deep SORT with low confidence track filtering

This version has the option to hide object detections instead of tracking. The settings in demo.py are

show_detections = TruewriteVideo_flag = TrueasyncVideo_flag = False

Setting show_detections = False will hide object detections and show the average detection confidence and the most commonly detected class for each track.

To modify the average detection threshold, go to deep_sort/tracker.py and change the adc_threshold argument on line 40. You can also change the number of steps that the detection confidence will be averaged over by changing n_init here.

Training your own models

YOLO v4

See https://github.com/Ma-Dan/keras-yolo4.

Deep SORT

Please note that the tracking model used here is only trained on tracking people, so you'd need to train a model yourself for tracking other objects.

See https://github.com/nwojke/cosine_metric_learning for more details on training your own tracking model.

For those that want to train their own vehicle tracking model, I've created a tool for converting the DETRAC dataset into a trainable format for cosine metric learning and can be found in my object tracking repository here. The tool was created using the earlier mentioned paper as reference with the same parameters.