Insect recognition and classification, Video surveillance

The proposed internship is within an Action of Collaborative Research between INRIA and its partners (Chambre d'agriculture des Alpes Maritimes, Institut National de Recherche en Agronomie, etc...) in the BIOSERRE project. The project aims at the design and implementation of a real-time surveillance system that can be used inside actual agricultural sites to allow early detection of infestations in cultures. The benefits from the usage of such surveillance systems can be huge both from an economic and an ecological point of view by

• allowing cost-effective intervention to fight against infestations
• minimizing losses in cultures caused by harmful insects
• favouring ecological solutions through introduction of auxilliaries (i.e. other insects known not be not harmful to the culture and to attack the eggs or the larvae of the harmful ones) thereby limiting chemical interventions (known to be harmful to the environment)

Moreover, the system is expected to be used by our partners to carry out biological experiments aiming at understanding insect behaviours.

Currenlty, an experimental greenhouse planted with roses and endowed with a network of Wifi cameras is beeing experienced in the framework of the BIOSERRE project. The cameras are connected to a standard PC. On the PC is implemented a video acquisition tool that allows to acquire video data in real time and at user-specified periods and to store these data on disk. Regarding the image processing part of our surveillance tool, a consequent work has been done regarding the detection of insects. However, to achieve a fully operative system to be used in actual agricultural sites, some aspects of the image processing module are still to be developed. These mainly concern intelligent motion detection --to optimize the camera network performance--, classification of detected objets, and their tracking between frames.

Indeed, the efficiency of the surveillance tool relies heavily on how it is able to classify each object detected by the detection system into the appropriate class:

• Actual insect or false alarm
• Propose a formal model of scenarios supporting suitable recognition algorithms.
• Harmful or unharmful insect
• For a harmful insect, recognize its species

This classification part is of key importance for our surveillance tool since it will allow to yield an efficient estimation of the number of harmful insects of each species in the field of view of the cameras so that an alarm of infestation can be declared and a rapid solution can be taken on time.

So, we expect the successful candidate to work on this classification part by designing and implementing efficient classification algorithms that can be used by our surveillance tool. As the tool is aimed to be used in an agricultural environment (e.g. a greenhouse), the classification algorithms must be robust to the presence of external agents (like humans working in the culture, animals (butterflies, frogs, etc...)), illumination changes, light reflections, limited image resolution, etc ... We shall consider the classification problem both from a static and dynamic point of view depending on whether or not time (thereby motion) is considered as a feature of the classication.

The candidate will experiment known state of the art classification techniques,(e.g. SVM classification, model selection) and evaluate their respective performances for the problem of classification of insects in videos. The most performant techniques found for the application will be implemented by the candidate in C/C++ and integrated into the image processing module of our video-surveillance tool.

The successful candidate will be a final year engineer student or a Masters student in computer science and/or applied mathematics, should have good knowledge in image processing and/or statistical learning, and should be familiar with C/C++ programming.

PULSAR Team, INRIA-Sophia Antipolis.