| Titre : | Real Time Multi-Object Tracking Using Deep Learning |
| Auteurs : | ABDELAZIZ HAMADI, Auteur ; Abdelmoumen Zerari, Directeur de thèse |
| Type de document : | Monographie imprimée |
| Editeur : | Biskra [Algérie] : Faculté des Sciences Exactes et des Sciences de la Nature et de la Vie, Université Mohamed Khider, 2021 |
| Format : | 1 vol. (121 p.) / ill. / 29 cm |
| Langues: | Anglais |
| Résumé : | The Deep Learning techniques has proven its effectiveness in so many computer vision (CV) tasks and one of those are the multi object detection and multi object tracking. Which comes in a lot of shapes and forms in terms of precision and speed.This master thesis goal falls under the Online Tracking detection-based method, to be able to manage Real-time Live scenarios where precision and speed are required for a quick and accurate response if needed. Where in our work, we only count on the current frames and past information (No peaking to future frames) to predict trajectories and track objects, which in our case is the “Pedestrian class”, we will use state-of-the-art and novel CNN based techniques in the object detection phase and in object tracking phase.We propose a better approach for an online Pedestrian tracking algorithm that is a robust and fast for an online application use, by reimplementing state-of-the-art algorithm DeepSORT, but instead of using its default detector Faster R-CNN, we replace it with a better state-of-the-art multi-object detector called YOLOv4.Since YOLOv4 didn’t give satisfactory results in terms of pedestrian detecting class specially in situations where there’s low light, extremely low light condition, or greyscale inputs, which means less information to process. And these kinds of situation are very important in terms of Pedestrian tracking and for security measurement, that’s duo to the fact that YOLOv4 was designed to be a good multi-class object detector where its goal to localize different object classes (80 class) as much as possible from a given scene.So, we customize the Object detector for our needs which we want a high precision in Pedestrian localization specially in low light and extremely low light conditions, with keeping the speed factor as important as the Precision, by customizing the architecture to focus on the localization of pedestrian’s class only, and retrain the model with a custom Pedestrian specified dataset. With the transfer learning techniques.After the preparation phase we reimplement the model with the new obtained weights to a TensorFlow format so it can be used in a GPU environment for faster object detecting calculation, to feed DeepSORT algorithm with good detections as much and as fast as possible.Then we build a bridge between the outputs of our model “Pedestrian_YOLOv4” and the inputs of DeepSORT algorithm so it can perform data association between the outputted detections in every frame by our model.Our experimental evaluation demonstrates that our techniques can produce high number of tracks with good overall precision, and an improvement in terms of speed specially in greyscale inputs and extremely low light conditions. |
| Sommaire : |
General Introduction ................................................................................... 14
Multiple Object Tracking in Computer Vision ............................................................. 14 Deep Learning in Object Detection ............................................. 14 Objective ................................................................... 15 Experiment multiple object detector algorithms using deep learning ...................................... 15 Chapter 1: Object Tracking .................................................. 17 Introduction ....................................................................... 17 1.1– Different Tracking Algorithms ...................................................... 17 1.1.1- Tracking Using Matching ............................................................ 17 1.1.2 – Tracking by detection: ......................................................... 17 1.2 – Brief Background on DEEP LEARNING (DL) ........................................................... 18 1.2.1- Machine learning overview (ML)........................................................................ 18 1.2.1.1 – Supervised Learning ......................................................... 19 1.2.1.2 – Unsupervised Learning: .......................................................................... 20 1.2.1.3 - Reinforcement learning: ........................................................................... 20 1.2.2 – Artificial Neural Network (ANN) ...................................................................... 20 1.2.3 - Convolutional Neural Network (CNN) ...................................................... 21 1.3.1 – Datasets ............................................................................. 23 1.3.1.1 – Object detection datasets ...................................................... 23 1.3.1.2 – Object tracking datasets ..................................................... 24 1.3.2 – Metrics ............................................................... 26 1.3.2.1 – Object detection metrics ....................................................... 26 1.3.2.1.1 - Intersection over Union (loU) ........................................................... 26 1.3.2.1.2 – Common Terms in Object Detection .............................................. 27 1.3.2.1.3– Average Precision (AP) ............................................................... 27 1.3.2.2 – Object Tracking metrics ..................................................................... 31 1.3.2.2.1 - Classical metrics ....................................................................... 31 1.3.2.2.2– CLEAR MOT metrics ............................................................................ 32 1.3.2.2.3– Identification (ID) metrics ........................................................... 33 Chapter 2 : Tracking-by-Detection using Deep Learning .......................... 35 Introduction ................................................................ 35 2.1- Object detectors ................................................................. 35 2.1.1- Backbone networks ....................................................................... 36 2.1.2 – Two stage Detectors ........................................................ 37 2.1.2.1 – R-CNN ................................................................................. 37 2.1.2.2 – Fast R-CNN .......................................................................... 38 2.1.2.3 – Faster R-CNN ......................................................................... 39 2.1.3– One stage Detectors .......................................................................... 40 2.1.3.1 – YOLO ................................................................................... 40 2.1.3.2 – YOLOv4 ........................................................................ 41 2.2 - MOT Trackers ........................................................................................... 46 2.2.1-MOT Categorization ................................................................ 46 2.2.2 - Online tracking algorithms ............................................................. 49 2.2.2.1 – Simple Online Realtime tracker (SORT) ...................................... 49 2.2.2.1.1– Detection ...................................................... 49 2.2.2.1.2– Estimation Model ............................................................... 50 2.2.2.1.3 – Data Association ......................................................................... 51 2.2.2.1.4 – Occlusion handling ..................................................................... 51 2.2.2.1.5 - Creation and Deletion of Track Identities .................................................. 51 2.2.2.2– Simple Online Realtime tracker with Deep associating metric (DeepSORT) .............. 51 2.2.2.2.1– Track managing & Estimation Model Update ............................................. 52 2.2.2.2.2 – Motion Estimation ....................................................................... 52 2.2.2.2.3 – Appearance features ....................................................................... 52 2.2.2.2.4– Cascade matching .......................................................................... 53 2.2.2.2.5 - Deep Appearance Descriptor ............................................................. 54 2.3 - Object Detection and Tracking Algorithms for Vehicle Counting ................................... 55 2.3.1 Methodology .................................................................................. 55 2.3.2- Used Algorithms for the comparative study ............................................ 56 2.3.3-Results ..................................................................... 56 2.4- Proposed Framework ............................................................. 59 Chapiter 3: Project Design ........................................................................... 63 Introduction ....................................................................... 63 3.1- System Design .............................................................. 63 3.1.1 Methodology .................................................................... 63 3.1.2 – Detailed System Design ............................................................... 65 3.1.2.1 – Pedestrian Detector Preparation Phase ....................................... 65 3.1.2.1.1– Dataset Collection ............................................................... 66 3.1.2.1.2- Dataset Reformatting ................................................................. 66 A. YOLOv4 Data Annotation Input .................................................................. 66 B. Dataset Annotation Reformatting ......................................................... 67 3.1.2.1.3– Training Pedestrian_YOLOv4 .......................................................... 68 3.1.2.1.4 – Data Augmentation ................................................................ 69 A. Self-Adversarial Training (SAT): ........................................................................ 69 B. Mosaic data augmentation: ............................................................................ 69 C. CutOut data augmentation: .............................................................................. 70 D. CutMix data augmentation: ............................................................................... 71 3.1.2.2 Application phase .................................................. 72 3.1.2.2.1 - Video Source Module .............................................................. 73 3.1.2.2.2 – Pedestrian Detection on GPU ............................................................ 73 3.1.2.2.3– Multi-Pedestrian Tracking with DeepSORT .................................... 74 3.1.2.2.3.1- Core Process ........................................................................... 74 3.1.2.2.3.2 – General Workflow .................................................................. 75 3.1.2.2.3.3 - Detailed Workflow .................................................................. 78 A. Kalman Filter Framework ........................................................................ 78 B. The Hungarian Algorithm (Kuhn-Munkres) ........................................ 83 C. IOU Match .................................................................................. 85 D. Cascade Matching ................................................................... 88 3.2– Project’s Workflow Summary ............................................................... 91 Chapter 4: Implementation and Results ....................................................... 93 4.1- Implementation ............................................................................... 93 4.1.1 – Environments and Developing tools ...................................................... 93 4.1.2-Environment details ......................................................... 96 4.2 – Results ....................................................................... 98 4.2.1 – Experiment Setup .................................................................................. 98 4.2.2 – FPS Results ................................................................... 99 4.2.3 – Results at Different Scenarios.......................................................... 99 4.2.3.1 – Walking on street (Camera moving) .............................................................. 99 4.2.3.2 – Football live Match ............................................... 100 4.2.3.3– Live Webcam .................................. 101 4.2.3.4 – Fixed Camera .................................................. 102 4.2.4 – Comparison in Night Time and Grayscale Cameras ........................................ 102 4.2.4.1 –Grayscale Security Camera. ................................................................... 103 4.2.4.2- Extremely Low Light Conditions (MOT challenge Video) .............................. 104 4.3 – limitations .............................................. 109 4.4 – Conclusion ............................................................ 110 General Conclusion ........................................................... 111 |
| Type de document : | Mémoire master |
Disponibilité (1)
| Cote | Support | Localisation | Statut |
|---|---|---|---|
| MINF/639 | Mémoire master | bibliothèque sciences exactes | Consultable |
