The animal-vehicle collision has been a topic of concern for years, especially in North America. To mitigate the problem, this thesis focuses on animal detection based on the onboard camera for intelligent vehicles. In the domain of image classification and object detection, the methods of shape matching and local feature crafting have reached the technical plateau for decades. The development of Convolutional Neural Network (CNN) brings a new breakthrough. The evolution of CNN architectures has dramatically improved the performance of image classification. Effective frameworks on object detection through CNN structures are thus boosted. Notably, the family of Region-based Convolutional Neural Networks (R-CNN) perform well by combining region proposal with CNN. In this thesis, we propose to apply a new region proposal method|Maximally Stable Extremal Regions (MSER) in Fast R-CNN to construct the animal detection framework. MSER algorithm detects stable regions which are invariant to scale, rotation and viewpoint changes. We generate regions of interest by dealing with the result of MSER algorithm in two ways: by enclosing all the pixels from the resulted pixel-list with a minimum enclosing rectangle (the PL MSER) and by fitting the resulted elliptical region to an approximate box (the EL MSER). We then preprocess the bounding boxes of PL MSER and EL MSER to improve the recall of detection. The preprocessing steps consist of filtering out undesirable regions by aspect ratio model, clustering bounding boxes to merge the overlapping regions, modifying and then enlarging the regions to cover the entire animal. We evaluate the two region proposal methods by the measurement of recall over IOU-threshold curve. The proposed MSER method can cover the expected regions better than Edge Boxes and Region Proposal Network (RPN) in Faster R-CNN. We apply the MSER region proposal method to the framework of R-CNN and Fast R-CNN. The experiments on the animal database with moose, deer, elk, and horses show that Fast R-CNN with MSER achieves better accuracy and faster speed than R-CNN with MSER. Concerning the two ways of MSER, the experimental results show that PL MSER is faster than EL MSER and EL MSER gains higher precision than PL MSER. Also, by altering the structure of network used in Fast R-CNN, we verify that network stacking more layers achieves higher accuracy and recall. In addition, we compare the Fast R-CNN framework using MSER region proposal with the state-of-the-art Faster R-CNN by evaluating the experimental results of on our animal database. Using the same CNN structure, the proposed Fast R-CNN with MSER gains a higher average accuracy of the animal detection 0.73, compared to 0.42 of Faster R-CNN. In terms of detection quality, the proposed Fast R-CNN with MSER achieves better IoU histogram than that of Faster R-CNN.

Thesis (MSc)--Stellenbosch University, 2015. / ENGLISH ABSTRACT: In South Africa, black-back jackal (BBJ) predation of sheep causes heavy losses to sheep farmers. Different control measures such as shooting, gin-traps and poisoning have been used to control the jackal population; however, these techniques also kill many harmless animals, as they fail to differentiate between BBJ and harmless animals. In this project, a system is implemented to detect black-backed jackal faces in images. The system was implemented using the Viola-Jones object detection algorithm. This algorithm was originally developed to detect human faces, but can also be used to detect a variety of other objects. The three important key features of the Viola-Jones algorithm are the representation of an image as a so-called ”integral image”, the use of the Adaboost boosting algorithm for feature selection, and the use of a cascade of classifiers to reduce false alarms. In this project, Python code has been developed to extract the Haar-features from BBJ images by acting as a classifier to distinguish between a BBJ and the background. Furthermore, the feature selection is done using the Asymboost instead of the Adaboost algorithm so as to achieve a high detection rate and low false positive rate. A cascade of strong classifiers is trained using a cascade learning algorithm. The inclusion of a special fifth feature Haar feature, adapted to the relative spacing of the jackal’s eyes, improves accuracy further. The final system detects 78% of the jackal faces, while only 0.006% of other image frames are wrongly identified as faces. / AFRIKAANSE OPSOMMING: Swartrugjakkalse veroorsaak swaar vee-verliese in Suid Afrika. Teenmaatreels soos jag, slagysters en vergiftiging word algemeen gebruik, maar is nie selektief genoeg nie en dood dus ook vele nie-teiken spesies. In hierdie projek is ’n stelsel ontwikkel om swartrugjakkals gesigte te vind op statiese beelde. Die Viola-Jones deteksie algoritme, aanvanklik ontwikkel vir die deteksie van mens-gesigte, is hiervoor gebruik. Drie sleutel-aspekte van hierdie algoritme is die voorstelling van ’n beeld deur middel van ’n sogenaamde integraalbeeld, die gebruik van die ”Adaboost” algoritme om gepaste kenmerke te selekteer, en die gebruik van ’n kaskade van klassifiseerders om vals-alarm tempos te verlaag. In hierdie projek is Python kode ontwikkel om die nuttigste ”Haar”-kenmerke vir die deteksie van dié jakkalse te onttrek. Eksperimente is gedoen om die nuttigheid van die ”Asymboost” algoritme met die van die ”Adaboost” algoritme te kontrasteer. ’n Kaskade van klassifiseerders is vir beide van hierdie tegnieke afgerig en vergelyk. Die resultate toon dat die kenmerke wat die ”Asymboost” algoritme oplewer, tot laer vals-alarm tempos lei. Die byvoeging van ’n spesiale vyfde tipe Haar-kenmerk, wat aangepas is by die relatiewe spasieëring van die jakkals se oë, verhoog die akkuraatheid verder. Die uiteindelike stelsel vind 78% van die gesigte terwyl slegs 0.006% ander beeld-raampies verkeerdelik as gesigte geklassifiseer word.

A research field currently advancing is the use of machine learning on camera trap data, yet few explore deep learning for camera traps to be run in real-time. A camera trap has the purpose to capture images of bypassing animals and is traditionally based only on motion detection. This work integrates machine learning on the edge device to also perform object detection. Related research is brought up and model tests are performed with a focus on the trade-off regarding inference speed and model accuracy. Transfer learning is used to utilize pre-trained models and thus reduce training time and the amount of training data. Four models with slightly different architecture are compared to evaluate which model performs best for the use case. The models tested are SSD MobileNet V2, SSD Inception V2, and SSDLite MobileNet V2, SSD MobileNet V2 quantized. Since the client-side usage of the model, the SSD MobileNet V2 was finally selected due to a satisfying trade-off between inference speed and accuracy. Even though it is less accurate in its detections, its ability to detect more images per second makes it outperform the more accurate Inception network in object tracking. A contribution of this work is a light-weight tracking solution using tubelet proposal. This work further discusses the open set recognition problem, where just a few object classes are of interest while many others are present. The subject of open set recognition influences data collection and evaluation tests, it is however left for further work to research how to integrate support for open set recognition in object detection models. The proposed system handles detection, classification, and tracking of animals in the African savannah, and has potential for real usage as it produces meaningful events

The Internet of Things is a constantly developing field. With advancements of algorithms for object detection and classification for images and videos, the possibilities of what can be made with small and cost efficient edge-devices are increasing. This work presents how camera traps and deep learning can be utilized for surveillance in remote environments, such as animal sanctuaries in the African Savannah. The camera traps connect to a smart surveillance network where images and sensor-data are analysed. The analysis can then be used to produce valuable information, such as the location of endangered animals or unauthorized humans, to park rangers working to protect the wildlife in these animal sanctuaries. Different motion detection algorithms are tested and evaluated based on related research within the subject. The work made in this thesis builds upon two previous theses made within Project Ngulia. The implemented surveillance system in this project consists of camera sensors, a database, a REST API, a classification service, a FTP-server and a web-dashboard for displaying sensor data and resulting images. A contribution of this work is an end-to-end smart surveillance system that can use different camera sources to produce valuable information to stakeholders. The camera software developed in this work is targeting the ESP32 based M5Stack Timer Camera and runs a motion detection algorithm based on Self-Organizing Maps. This improves the selection of data that is fed to the image classifier on the server. This thesis also contributes with an algorithm for doing iterative image classifications that handles the issues of objects taking up small parts of an image, making them harder to classify correctly.